Implementation of a breast cancer
screening prediction algorithm:
a knowledge management approach
A research project funded by the
NHS Cancer Screening Programmes
Research Team
Dr Rajeev K Bali
(Principal Investigator)
Dr Vikraman Baskaran
Dr Hisbel Arochena
Prof Raouf Naguib
Knowledge Management for Healthcare (KARMAH) research subgroup,
Biomedical Computing and Engineering Technologies (BIOCORE)
Applied Research Group,
COVENTRY UNIVERSITY,
Health Design and Technology Institute (HDTI),
Coventry University Technology Park, Puma Way,
COVENTRY, West Midlands CV1 2TT
www.coventry.ac.uk/biocore/
www.coventry.ac.uk/karmah/
September 2009
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Acknowledgements
The BIOCORE research team would like to thank Julietta Patnick CBE, Director, NHS
Cancer Screening Programmes for funding this research project. Additional thanks are due
to Dr Matthew Wallis (now with the Cambridge Breast Unit and the National Institute for
Health Research (NIHR) Cambridge Biomedical Research Centre) and Margot Wheaton
(Warwickshire, Solihull & Coventry Breast Screening Unit) for their support.
Disclaimer
This report is the edited output of a PhD research project carried out by the Biomedical
Computing and Engineering Technologies (BIOCORE) Applied Research Group at Coventry
University. It is provided for information only to interested practitioners and for academic
research purposes. For further information, please contact Dr RK Bali (r.bali@ieee.org).
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Abstract
The Warwickshire, Solihull and Coventry Breast Screening Unit (‘Unit’) is one of the largest
screening services in the UK. In collaboration with the Unit, an Artificial Intelligence Attendance
(AI-ATT) algorithm was recently developed within Biomedical Computing Research Group
(BIOCORE), for the prediction of attendance (starting from the second episode) to the Breast
Screening Programme (BSP). This research aims to integrate a new algorithm (entitled “Java-based
Attendance prediction by Artificial Intelligence for Breast Screening (JAABS)” within this Unit in
order to streamline and automate the invitation process and thus increase screening attendance. The
work deploys a hybrid research methodology which incorporates both the qualitative and quantitative
paradigms. Efforts had focussed on the formulation of a BSP messaging protocol. The protocol
processes the uptake dataset from the second episode upwards and uses it as the input for the
prediction process. The non-attendee list produced by the AI-ATT algorithm is segregated into their
respective GP groups. Using the proposed framework, a message transmission process will be
initiated to send the information to the GPs’ network. Following proper authentication and
acceptance, the information is updated on the GP’s database. This can instigate the deployment of
measures to improve attendance and potentially result in a further reduction in morbidity and
mortality rates due to breast cancer. Knowledge Management (KM) tools and techniques were used
to effectively manage the transfer of knowledge between the Unit, GPs and eligible women within a
screening episode. In concurrence with the UK government’s “e-Government Interoperability
Framework” (e-GIF) policy, and in projecting the importance of the Electronic Data Interchange
(EDI) within the BSP structure, this project addresses the limitations of the existing communication
path between the BSP and primary care providers. The deliverables of the project include an
investigation and evaluation of the suitability of HL7 and XML standards for the proposed messaging
protocol, a comparison of the proposed standards with EDIFACT-based protocols to highlight salient
features. As part of the research, questionnaires were designed and distributed to evaluate the GPs’
role in reducing the BSP non-attendance. A comprehensive triangulation-based validation of the two
algorithms (AI-ATT and JAABS) was also featured as part of the research. The JAABS algorithm
proposed includes a new attribute (associated to the distance travelled by the women for a
mammogram) as a predicting variable to increase the accuracy of prediction. Based on these
evaluations and validations, technical reports were produced and communicated to all stakeholders.
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Contents
List of Figures
List of Tables
Introduction
Setting the scene
Breast Screening Programme in the United Kingdom
Challenges being addressed
Solution proposed
Research Question
Research Objectives
Overview
1
1
3
6
7
8
9
12
Proposed Knowledge Exchange
Information and knowledge
Knowledge creation
Knowledge Management (KM)
Beyond KM
Total Knowledge Management in healthcare (TKMh)
Knowledge creation and application of TKMh model for this project
Knowledge sharing for this project
Proposed twin approach
Prediction and post-screening results
HL7 Message design
Modelling methodology
Information analysis
Information modelling
HL7 v3 Message implementation
Challenges faced during message design and implementation phase
Proposed protocol
Proposed Breast Screening Attendance Messaging Protocol (BSAMP)
Conclusion
14
15
15
16
20
23
24
28
29
30
32
33
33
35
36
37
39
39
42
Algorithm for Screening Attendance Prediction
Predecessor algorithm employed in exploratory research
Artificial Intelligence based Attendance Prediction algorithm (AI-ATT)
Analysis of AI-ATT
AI-ATT working environment
Data pre-processing
Data Pre-processing design
Module design
Variables identified for data pre-processing
Data pre-processing module implementation
Proposed algorithm
Java based Attendance prediction by Artificial Intelligence for
Breast Screening (JAABS) algorithm
JAABS design
JAABS classes
Testing JAABS algorithm with simulated dataset
Comparison of AI-ATT with JAABS implementation
JAABS evaluation
Dataset
43
43
44
44
45
48
48
49
50
52
53
53
53
55
56
59
60
60
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Attribute selection
JAABS vs AI-ATT attributes
Screening distance as a new predictor attribute
JAABS ROC tests
RBFN test
ANNP test
JAABS test
Summary
65
66
67
68
68
71
73
75
Knowledge Sharing: GP’s Role
GP intervention
Types of intervention
Questionnaire-based survey
Questionnaire analysis
Screening women
Prediction
Electronic Data Interchange (EDI)
GP intervention
Additional information on interventions
Questionnaire results
Survey summary
76
76
76
77
79
80
81
81
82
82
82
87
Conclusion and Future Work
Conceptual development
Concept development - Knowledge management
Concept development - Artificial Intelligence (AI)
Concept development - Health Informatics
Original contribution
Objectives accomplished
Project deliverables accomplishment list
Adopting the proposed protocol
Challenges
Future work
Concluding observations
88
89
89
89
90
90
92
93
93
94
96
97
References
100
Appendix: Research Papers Published
113
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
List of Figures
FIGURE 1 - BREAST ANATOMY © (NATIONAL CANCER INSTITUTE 2005)
FIGURE 2 – MAMMOGRAPHY MACHINE AND TWO VIEWS OF MAMMOGRAM
FIGURE 3 - BREAST SCREENING UPTAKE FROM 1995-2005
FIGURE 4 - CANCER DETECTION IN UK FROM 1995 TO 2005
FIGURE 5 - CANCER DEATHS REPORTED IN UK FOR 1995 – 2004
FIGURE 6 - TYPICAL WEB SERVICES IMPLEMENTATION FOR THE PROPOSED PROTOTYPE
FIGURE 7 – CHAPTERS’ RELATIONSHIP WITH THE RESEARCH CORE AREAS
FIGURE 8 - KNOWLEDGE CREATION CYCLE (WITH 1ST AND 2ND GENERATION INFORMATION)
FIGURE 20 - KNOWLEDGE CREATION SPIRAL
FIGURE 21 - MULTIPLE FACETS OF KNOWLEDGE MODEL IN PROJECT-BASED ORGANISATION
FIGURE 22 – ORGANISATIONAL KNOWLEDGE SHARING (INTER AND INTRA MODES)
FIGURE 23 - TKMH FRAMEWORK
FIGURE 24 - KNOWLEDGE CREATION IN PREDICTING SCREENING ATTENDANCE
FIGURE 25 - TKMH FRAME WORK APPLIED FOR CURRENT RESEARCH
FIGURE 26 - SCREENING ATTENDANCE PREDICTION THROUGH ANN
FIGURE 27 - KM-BASED ARCHITECTURE FOR INCREASING SCREENING ATTENDANCE
FIGURE 28 - INCREASING SCREENING ATTENDANCE-TWIN APPROACH
FIGURE 29 - CLASS DIAGRAM FOR SCREENING SCENARIO
FIGURE 30 - DESIGN OF MESSAGE (SCREENING) WITH OBJECTS AND RELATIONSHIPS
FIGURE 31 - SAMPLE SCREENING REPORT SENT TO GP IN HARD-COPY FORMAT
FIGURE 32 - INTERACTION DIAGRAM FOR THE TWIN MESSAGES
FIGURE 33 - BSAMP PROTOCOL ARCHITECTURE
FIGURE 34 - SCREEN SHOT OF PREDICTION AND NON-ATTENDANCE DISPLAYED IN FRAMES WITH
DIFFERENT SET OF FONT COLOURS
FIGURE 35 – SCREEN SHOTS OF PREDICTED NON-ATTENDANCE DISPLAYED AS ‘AMBER’ AND
NON-ATTENDANCE AS A ‘RED’ POP-UP PROMPT ON THE WOMEN’S CLINICAL RECORD
FIGURE 36 – DATA MINING STREAM FOR AI-ATT ALGORITHM (ADAPTED FROM
AROCHENA’S PROJECT REPORT (AROCHENA 2003)
FIGURE 37 - AI-ATT ALGORITHM WORKING PROCESS (AROCHENA 2003)
FIGURE 38 - PRE-PROCESSING MODULE FOR PROPOSED PROTOTYPE
FIGURE 39 - HIGH LEVEL USE CASE DIAGRAM OF THE PROTOTYPE
FIGURE 40 – DATA PRE-PROCESSOR MODULE CLASS DIAGRAM
FIGURE 41 - AI MODULE (JAABS ALGORITHM IMPLEMENTATION)
FIGURE 42 - JAABS ALGORITHM (CLASS DIAGRAM)
FIGURE 43- JAABS ALGORITHM
FIGURE 44 - THE ROC CHARACTERISTICS FOR THE JAABS ALGORITHM
FIGURE 45 - ROC VALUES ON THE PREDICTED WOMEN (IN NUMBERS) FOR BPNN AND RBFN NETWORK
FIGURE 46 - TEST RUN FOR CR TEMPLATE VERIFICATION
FIGURE 47 - SELECTION CRITERIA USED FOR THE ENTIRE DATASET
FIGURE 48 - FINAL VERSION OF CR TEMPLATE
FIGURE 49 - PIE-CHART SHOWING RECORDS CLASSIFICATION AFTER PRE-PROCESSING
FIGURE 50 - TRI-FURCATION OF DATASET INTO TRAINING, VALIDATION AND TESTING SET
FOR SEVEN EPISODES
FIGURE 51 - RBFN STREAM PREDICTION DETAILS FOR ALL EPISODES
FIGURE 52 - PREDICTION CHARACTERISTICS FOR ANNP
FIGURE 53 - JAABS VS AI-ATT AVERAGE ROC VALUES FOR FIRST FOUR EPISODES
FIGURE 54 - CONSULTATION RATE PER PERSON YEAR TREND (1995-2006) FOR AGE GROUP
BETWEEN 50 AND 69
1
3
4
4
5
8
12
14
16
18
22
23
24
25
27
28
30
32
33
34
35
40
41
42
45
46
48
50
51
54
55
56
57
58
61
61
62
63
64
70
72
75
83
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
List of Tables
TABLE 1 - EXTRAPOLATION OF CANCER NOT DETECTED DUE TO NON-ATTENDANCE........................................................6
TABLE 4 - PRE-SCREENING INTERACTION TABLE (PRSC_IN060001).............................................................................................. 36
TABLE 5 - POST-SCREENING INTERACTION TABLE (PRSC_IN060002) ........................................................................................... 36
TABLE 6 – CRYSTAL REPORT TEMPLATE FOR DATA EXTRACTION .............................................................................................. 51
TABLE 7 - JAABS ALGORITHM'S ROC METRICS .................................................................................................................................... 57
TABLE 8 - RECORDS BREAKDOWN DETAILS AFTER DATA PRE-PROCESSOR MODULE ......................................................... 62
TABLE 9 - TRAINING, VALIDATION AND TEST SETS CLASSIFIED BY THE PRE-PROCESSOR MODULE ............................. 63
TABLE 10 - PREDICTOR ATTRIBUTES FOR AI-ATT ALGORITHM..................................................................................................... 65
TABLE 11 - VARIABLES ASSOCIATION LEVEL WITH SCREENING ATTENDANCE FOR THE NEW DATASET................... 66
TABLE 12 - AVERAGE ROC CHARACTERISTICS FOR JAABS WITH AI-ATT VARIABLES........................................................... 67
TABLE 13 - SCREENING DISTANCE TEST ROC VALUES ..................................................................................................................... 67
TABLE 14 - ROC CHARACTERISTICS FOR JAABS ALGORITHM ........................................................................................................ 73
TABLE 15 - ROC CHARACTERISTICS OF AI-ATT ALGORITHM ......................................................................................................... 74
TABLE 16-ISSUES COVERED BY THE QUESTIONNAIRE (BASKARAN ET AL. 2006) .................................................................... 78
TABLE 17- DICHOTOMISATION OF THE VARIABLES CATEGORISED BY QUESTIONNAIRE SUBSECTIONS...................... 79
TABLE 18 - CROSS TAB ANALYSES ON SCREENING, PREDICTION, EDI AND INTERVENTION AGAINST
DEMOGRAPHIC DETAILS .................................................................................................................................................................. 80
TABLE 19 - PROJECT WORK'S ORIGINALITY SUMMARY.................................................................................................................... 92
TABLE 20 - PROBABLE NUMBER OF LIFE SAVED BY JAABS ALGORITHM.................................................................................... 93
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Introduction
We provide an introduction to the project work and its focus, objectives and deliverables.
This project proposes a novel approach for leveraging knowledge through prediction in a
multi-tier architecture. The knowledge leveraging addresses the breast screening nonattendance and its related issues. This project work is aimed at providing the first step
towards a multi-pronged strategy envisaged to deliver better primary and secondary care
from breast screening perspective. This strategy is also the first of its kind in the UK that
addresses breast screening non-attendance from a Knowledge Management perspective.
Setting the scene
Breast cancer is the most common cancer in women with over forty thousand women being
diagnosed with the disease each year in the UK (Cancer Research UK 2005). Any
information related to the breast can largely affect a women’s consciousness and a threat of
breast cancer will have varying impacts on women’s psychology (Cassandra 2006). The
incidence of breast cancer is greater within higher social classes resulting in increased
awareness of breast cancer and its prevention (Chris et al. 2003). Typically breast cancerous
cells originate in the mammary glands (lobules) or in the ducts connected to these glands
(Figure 1) or in other tissues around these glands (American Cancer Society 2005). When in
close proximity to the lymphatic system, these cells can result in being carried to other
organs of the body. This subsequently results in cancerous growth in that organ and is
described as metastatic breast cancer (American Cancer Society 2005). Although many
causes had been identified of breast cancer, the knowledge of finding a cure is still not
within the reach of modern medicine.
Figure 1 - Breast anatomy © (National Cancer Institute 2005)
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
1
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Breast cancer should ideally be diagnosed at the earlier stages of its development. Possible
treatments include removing or destroying the cancer cells to avoid the spread of the
affected cells. Breast self examination (BSE) is an effective and non-intrusive type of self
diagnosis exercise for checking any abnormalities/lumps in the breast tissue. Unfortunately
this greatly depends on the size of the lump, technique and experience in carrying out a self
examination by the woman (Oikonomou et al. 2003). An ultrasound test using sound waves
can be used to detect lumps but this is usually suited for women aged below thirty-five
owing to the higher density of breast tissue (American Cancer Society 2005).. Having a
tissue biopsy via a fine needle aspiration or an excision is often used to test the cells for
cancer. These tests are mostly employed in treatments or post-treatment examination and as
second rung diagnostic confirmation methods (Marcela 2004). Performing a Computed
Tomography (CT) or a Magnetic Resonance Imaging (MRI) scan would result in a thorough
examination of the breast tissue but this technique is not favoured due to reasons including
that it may not be economical, needs preparation, is noisy, time consuming and the images
may not be clear (Marcela 2004).
Mammography is a technique for detecting breast tissue lumps using a low dosage of X-ray.
This technique can even detect a three millimetre sized lump. The X-ray image of the breast
tissue is captured and the image is thoroughly read by experienced radiologists and
specialist mammogram readers (Marcela 2004). Preliminary research suggests that women
aged fifty five and above are more susceptible to breast cancer; mammography is more
suited to the women aged fifty five and above (due to the lower density of breast tissue)
(Blanks et al. 2000).. Even though mammography has its critics - mainly due to its high rate
of false positives and false negatives (Epstein 1979, Burton 1997) - it has still become the
standard procedure for screening women by the National Health Service (NHS) National
Breast Screening Programme UK (Forrest 1986). Mammography is the best and most viable
tool for mass screening to detect cancer in the breast at an early stage (Medicine net 2002)
however, the effectiveness of diagnosis through screening is directly dependent on the
percentage of women attending the screening programme (Pirjo et al. 2001).
The NHS Breast Screening Programme, catering to the entire eligible female population is
funded by the Department of Health in the UK and is the first of its kind in the world. It
covers nearly four million women and detected more than thirteen thousand cancers in the
screened population for the year 2005 (NHS review 2006). Currently the screening
programme routinely screens women between the ages of fifty and seventy, and employs
two views (Figure 2) of the breast, medio-lateral and cranio-caudal (Department of
radiology, Brigham and women’s hospital 2007).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
2
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Craniocaudal
Mediolateral
Figure 2 – Mammography machine and two views of mammogram
Breast Screening Programme in the United Kingdom
High mortality and morbidity rates due to breast cancer have sustained the need for a
continuous impetus on research and strategies focussed on preventive medicine (Turner et
al. 1994, Majeed et al. 1997, Sin, and Leger 1999). Breast cancer is still one of the major cancer
killers in women. The lack of breakthroughs in finding a definitive cure means that
preventive medicine is the only viable alternative in reducing deaths due to breast cancer
(Fox et al. 1991).. Various randomised controlled trials (RCTs) in Europe and around the
world have concluded that screening can be an effective tool in fighting breast cancer
(Bjurstam et al. 1997, Majeed et al. 1997, Duffy et al. 2003). Technology breakthroughs in
imaging and digitization have increased the efficacy of mammograms and countries such as
the UK, Sweden, Australia, Canada and the USA all have well-established breast screening
programmes (Turner et al. 1994, Majeed et al. 1997, Sin and Leger 1999, Taylor et al. 2003).
The UK National Health Service (NHS) National Breast Screening Programme (NBSP) is
unique as it provides free breast screening for the female population aged between 50 and 70
at a national level (Cancer Research UK 2004, Forrest 1986, Tabar et al. 1985).
The recent increase of the upper age limit from 63 to 70 for screening and making a twoview mammogram mandatory has greatly increased the efficiency of benign or malignant
tumour detection (Patnick 2006). The NBSP currently runs a massive screening programme
catering to almost two million eligible women across the UK (Cancer Research UK 2004).
This programme runs on a call/recall cycle which screens all eligible women in a three year
interval. The information published by the UK Government Statistical Service (NHS Health
and Social Care Information Centre) in its Community Health Statistics report for the year
2006 agrees that, for the past ten years since 1995, the uptake has remained constant at
around 75% (Protti 2005) (Figure 3).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
3
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
80
Screening uptake
78
76
74
%
72
70
68
66
64
97
1995 96
98
Women aged 50-64
99
2000
01
02
04
03
Women aged 65-70
2005
Figure 3 - Breast screening uptake from 1995-2005
From 2002, women aged 65 to 70 had an uptake of 70%. In addition to this, the report also
indicated that the number of cancers detected steadily increased from over 5000 in 1995 to
nearly 12000 in 2005 – refer Figure 4 (NHSBSP 1999-2005). Deaths due to cancer have
gradually decreased from over 11,000 in 1995 to approximately 10,000 (Figure 5). This not
only suggests that the number of clinically detected breast cancers has increased, but also
confirms that deaths due to breast cancer have steadily declined over the years (using the
data until 2004, NHSBSP 1999-2005).
14
Women with cancer detected
12
Thousands
10
8
6
4
2
0
1995
96
97
98
99
2000
01
02
03
04
2005
Figure 4 - Cancer detection in UK from 1995 to 2005
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
4
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
12
Cancer Deaths
Thousands
11.5
11
10.5
10
9.5
1995
96
97
98
99
2000
01
02
03
2004
Figure 5 - Cancer deaths reported in UK for 1995 – 2004
All these inferences indicate that mammogram-based screening initiatives are valuable and
do contribute to reduce breast cancer related deaths. The question is whether the screening
programme is working efficiently. The efficiency can be mapped to the screening
attendance. Figure 3, Figure 4 and Figure 5 clearly reveals that the most important inference
that can be deduced from these figures is that the number of non-attendees has been
significantly increasing and has reached half a million.
Simple projection of this data, through translation (Table 1), submits that nearly 4,000 cancer
incidences would not have been diagnosed. Even if a small percentage of these nonattendees could be made to attend, it would result in the saving of significant lives.
Indirectly it can also be infered that, despite focussed efforts on these non-attendees for the
past ten years, there was no real effect on their attendance. Moreover, early stage cancer
detection would have a huge impact in reducing cancer related deaths (Baskaran et al. 2006a,
2006b). From these facts and data, it can be confirmed that the primary concern is to reduce
non-attendance (Bankhead et al. 2001).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
5
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Screening
Year
Number of
women
invited
Acceptance
rate
Number of
cancers
detected
Probable number
of cancers not
detected
1996-97
1,558,995
75.1
7,141
1,776
1997-98
1,668,476
75.5
7,932
2,635
1998-99
1,669,727
75.5
8,771
2,580
1999-00
1,811,541
75.4
9,525
3,110
2000-01
1,815,610
75
9,866
3,286
2001-02
1,752,526
75.5
10,003
3,237
2002-03
1,873,470
74.7
10,467
3,539
2003-04
1,998,989
75.2
13,064
4,284
Table 1 - Extrapolation of cancer not detected due to non-attendance
Challenges being addressed
Detecting cancer at its early stages is the only viable strategy to decrease cancer related
mortality. This decisive aspect has driven the urgent need to focus more on screening
attendance. Maximising screening attendance can only be achieved through a well
coordinated delivery of better healthcare (healthcare) to the population. Based on reports by
two Swedish county trials (Bjurstam et al. 1997, Duffy et al. 2003) and subsequent
recommendations by Forest in 1986, a 70% attendance was proposed as a benchmark for
future coverage statistics. In spite of various efforts focussed on this issue, the attendance
rate has not improved since this benchmark metric. Even though the 70% screening
attendance initially seemed to be a challenging statistic to be achieved, it was never
improved upon over the years. The current UK national attendance target for breast
screening is 80%. The increased screening population can be attributed to a variety of
reasons including the post second world war population explosion, “baby boomers” and the
general longevity of human life in this century that has forced health strategists to rethink
existing health policies. This has created the opportunity for continuous research, aimed at
different facets of care delivery. One such area was the age limit prescribed for the breast
screening programme.
Such continuous research had indicated that it is a viable strategy to increase the upper age
limit from 64 to 70, thus providing extended mandatory care delivery periods (DOH 2006).
According to the Warwickshire, Solihull and Coventry Breast Screening (WSCBS) service
manager; the National Screening Service is seriously considering to increase the age band at
both ends by an additional three years (breast screening process would start when women
are aged 47 years and continue until they are 73 years). This would effectively result in the
increase of the eligible women population who would be participating in the future NBSP.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
6
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
As indicated earlier, the necessity for increasing the quality of care services has prompted
the NHS Cancer Screening Programmes to offer a “two view” mammogram (in 2003) as a
way to improve the efficacy of detecting screening abnormalities (Patnick 2004, Osborna et
al. 2006). Coupled to this, the initiative to gradually digitise mammograms had called for
increased investments and has motivated screening units across the country to take
advantage of the new technologies available. The digitization of mammograms has not only
improved the quality of images but has also facilitated the access, storage (archiving) and
sharing of images across geographical boundaries enabling better utilisation of image
readers’ (Pisano et al. 2002, Skaane et al. 2003, Skaane and Skjennald 2004).. Such mentioned
care enhancements are putting undue pressure on the NHS Cancer Screening Programme
which is already overstretched in terms of resources. The problems became more complex
when such resources are underutilised due to static, below-par attendance rate. Proactive
and acute healthcare are vital for cancer control and cure; this has reinforced the necessity to
improve screening attendance.
Solution proposed
The above-mentioned challenges can only be addressed by a resource-saving strategy which
has better healthcare at its core. As part of this strategy Knowledge Management (KM) was
identified as the holistic tool. Techniques associated with knowledge creation and sharing
were viewed as approaches for better resource utilisation. Based on the above specifics, an
automated software solution - coupled with a humanistic strategy - was designed. The first
component in the proposed strategy is related to knowledge creation via Artificial
Intelligence (AI), employing Neural Network (NN) algorithms. This strategy also includes a
Service Oriented Architecture (SOA) to deliver the envisaged knowledge sharing as the
second component.
This research proposes to unify the existing National Breast Screening Computer System
(NBSS) software onto a single platform and create prototype software components based on
Open Source technologies. The proposed prototype software would be automated to
produce the pre-processed data and eventually normalise the data for AI (neural network)
assimilation. These activities would be performed sequentially without human involvement
for repeatability, reliability and accuracy. The AI model for attendance prediction itself
would be simulated on the Open Source technology platform. This model incorporates all
additional transformations occurring within the screening process (including the change in
the screening upper age limit).. The prototype framework proposed will incorporate the AI
model for knowledge creation (i.e. list of predicted non-attending women). The prototype
combines the demographic data pertaining to the non-attending women and information
related to her General Physician (GP) as a messaging package. This package triggers the
generation of an electronic message based on the Health Level 7 (HL7) version 3 standards
and utilises Service Oriented Architecture (SOA) as the message delivering technology (refer
to Figure 6).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
7
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Services provider
Services directory
Define service
& advertise
Services offered
Like lab results
through NHS SPINE
Services contract
accepted
Services requested
after login into NHS
SPINE
Search for
services
Client (service requestor like BSU, GP)
PC
Laptop
Workstation
PDA
Mobile
Figure 6 - Typical Web Services implementation for the proposed prototype
Research Question
Earlier research in identifying the factors contributing to non-attendance prediction was
tested in the WSCBS unit (Arochena 2003). This research was the first step in identifying
attributes that can predict women’s attendance a priori. This prediction algorithm was
developed by an expert statistician through visual modelling software in an academic
setting and there is a lack of associated research as to how this prediction can be put to
effective use. Over the last decade, international research pursued by Melville et al. 1993,
Bryant 1996, Richards et al. 2001 and Zorbas 2003 had indicated without doubt that primary
care can play a vital role in alleviating the problem of non-attendance through interventions
(Bankhead et al. 2001, Brunton and Thomas 2002).
My primary research question is:
Can the breast screening attendance be increased through appropriate
Knowledge Management tools and techniques?
The hypotheses this project sets out to validate are:
•
that it is possible to design and implement an Open Source-based prediction
algorithm prototype for breast screening attendance
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
8
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
•
•
that it is possible to design a messaging platform for knowledge creation
(through prediction) and sharing model (knowledge/information exchange
architecture) with available technology and infrastructure
that it is feasible for GPs to make interventions to increase breast screening
attendance through electronic prompts
Research Objectives
This project proposes to integrate the existing call / recall process of the Breast Screening
Programme (BSP) with the results obtained by a newly developed predictive algorithm in
order to increase uptake in the screening programme. This will be achieved through the use
of Knowledge management (KM)-based tools and techniques. The developed protocol will
be the first step in providing the ability to target non-attendance at an early stage and
initiate counter measures automatically.
The objectives included:
•
•
•
•
•
To investige and evaluate current practices in the BSP call / recall process
To investigate and evaluate the suitability of the existing Electronic Data
Interchange For Administration, Commerce and Transport (EDIFACT) and
Health Level 7 (HL7) standards
To design and analye a questionnaire to evaluate the GPs’ role in reducing
BSP non-attendance
To identifify a messaging platform and a viable architecture for integrating
the Java based Attendance prediction by Artificial Intelligence for Breast
Screening (JAABS) algorithm
To develop a Knowledge Management protocol based on the above, followed
by the implementation of a prototype (JAABS algorithm) encompassing the
required standard, ready for knowledge sharing with healthcare stakeholders
via the NHS’ National Care Records Services (NCRS) infrastructure.
The primary deliverables of this work are a set of breast screening messaging protocols
which meet the standards set by Health Level 7 (HL7) UK and implemented via tools
recommended by HL7 and the National Health Service (NHS) Connecting for Health (CfH).
The work detailed in this project builds on an earlier research conducted within the
Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group
at Coventry University. The information and knowledge sharing framework that has been
proposed for deployment on a Service Oriented Architecture (SOA) operates on web
services-based technology and draws parallel to the UK government’s e-gif policy (Ward
2003). The prototype proposed would be the first software iteration in building a real-time
functioning KM-based system which would enable a seamless transfer of knowledge for
better healthcare delivery involving screening unit and primary care (GP).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
9
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
10
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The project deliverables are:
•
•
•
•
an evaluation report on the suitability of the HL7 Version 3.0-based XML
standards would be conducted for its appropriateness to the proposed BSP
Messaging (BSPM) protocol
an extension of the evaluation report to compare the proposed HL7 Version
3.0 with EDIFACT based protocol for highlighting salient features
a framework for implementing the JAABS algorithm
a prototype implementation to predict non-attendance by the JAABS
algorithm
•
tests of the execution, evaluation of the prototype with real-time dataset
•
a design for the proposed BSPM messaging protocol will be designed
•
an architecture for the proposed prototype
•
•
•
a questionnaire-based evaluation of the GPs’ role in reducing the BSP nonattendance and justification of the proposed prototype’s KM based strategy
technical reports
dissemination of findings would be via academic presentations, paper
publications and conferences
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
11
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Overview
The current research can be visualised as a complete set of proposed protocols within the
healthcare domain. The core component, namely Knowledge Management (KM), Artificial
Intelligence (AI), and Health Informatics (HI), brings the necessary synergy into the research
and transforms it into truly inter-disciplinary work (Figure 7). This invigorates the breast
screening domain and provides the impetus required for new approaches.
CH2
CH4
CH2
CH4
CH1
CH6
CH5
CH2
CH3
Figure 7 – Chapters’ relationship with the research core areas
The specific components of the current research and their relationship to the research
chapters are shown in Figure 7. The diagram appropriately positions the chapter with
reference to their content and its association to the core areas of the research. Some chapters
are repeated in different parts of the diagram; that signifies a multidisciplinary coverage of
the specific chapter.
We discuss the introduction, challenges being addressed and the research objectives.
Chapter 2 (which has been removed for the purposes of this document) details the literature
review and a summary of the research areas related to this study namely Knowledge
Management, Health Informatics and Artificial Intelligence. As part of the messaging
standard evaluation, a comparison between EDIFACT and HL7 is included at the end of this
chapter. We provide the details of knowledge, its creation and sharing. This is followed by a
framework which highlights the human-centric nature of knowledge and a methodology
based on KM (specifically for healthcare domain) is proposed. This chapter concludes with a
detailed description of the proposed Breast Screening Attendance Messaging Protocol
(BSAMP) and substantiates the importance of HI in healthcare knowledge and information
dissemination.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
12
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
We go on to deal with the AI component of this project which incorporates data processing
and the proposed JAABS algorithm and its evaluation with the earlier algorithm Artificial
Intelligence based ATTendance predicting algorithm (AI-ATT). This chapter also
enumerates the design and implementation logic employed by the JAABS algorithm. An
evaluation and testing report of the JAABS algorithm concludes this chapter. Chapter 5
identifies the GPs’ role for the knowledge assimilation and the types of intervention that can
be initiated to increase the screening attendance. A comprehensive analysis of the
questionnaire on GP intervention and results are also discussed in this chapter.
We conclude this project and highlights the changes necessary for adopting the proposed
BSAMP protocol. This project also lists the conceptual developments accomplished in a
separate subsection and highlights how the objectives were not only achieved but, in many
cases, were exceeded. This chapter shows the full richness of this project and reflects the
new concepts added within the fields of healthcare informatics and Knowledge
Management and how these new concepts will further the knowledge boundaries. The last
subsection details the challenge faced and lays out future work for continued research in this
domain.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
13
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Proposed Knowledge Exchange
This research illustrates the importance of Knowledge Management (KM) and its effective
usage in healthcare domain. Newcomers to KM, especially healthcare specialists, are often
overwhelmed when trying to understand the intricacies of KM in healthcare. These
challenges provided the necessary impetus to the author for investigating the various KM
concepts and elucidate its principle components with clarity for easier understanding.
Figure 8 - Knowledge creation cycle (with 1st and 2nd generation information)
The core for KM is knowledge and its management of its various avatars. It is fundamentally
important to understand and define knowledge from the KM perspective. The following
sections deals with the various views and suggestions from the author on KM, starting with
a definition of knowledge from KM context.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
14
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Information and knowledge
A misconception exists in differentiating knowledge from information that has attracted a
continued debate on the legitimacy of the KM paradigm itself (Wilson 2002). The schematic
depicted in Figure 8 aims at differentiating information and knowledge within a healthcare
perspective. The figure's analogy of data can be interpreted as a set of raw, unprocessed
numbers (75, 77, 57, 70, 72); contextual details when added to this set of data - such as (1)
this dataset represents heart beats per minute and (2) pertaining to male aged 62 years -can
add lucidity to the dataset provided, becoming information (Miller 2002).. When a heart
specialist assimilates this information, a unique process (which is an inherent characteristic
of the human brain) of identifying connections, similarities and patterns in the information
provided using his/her earlier insight on the subject initiate new knowledge about the
patient and may result in the conclusion of symptoms of Arrhythmia (Koskinen 2003). The
said steps constitute the initial phase of the knowledge cycle.
Such knowledge cycles will contribute further knowledge creation and result in adding
clarity and assist in achieving wisdom. As a routine process of sharing this acquired
knowledge, an individual has to resolve this issue through traditional knowledge sharing
techniques such as writing notes, manuals, reports and books coupled with lectures,
meetings and the like. The knowledge residing in the human brain is termed Tacit (or
implicit) knowledge and the other externalised forms of it are termed Explicit (or expressed)
knowledge. There is a grey area in differentiating the explicit knowledge from information.
Under some circumstances, these can become interchangeable but the human context alone
differentiates the information from explicit knowledge (Augier et al. 2001). In expressing
his/her tacit knowledge, the heart specialist resorts to the creation of notes based on the
patient's information. This form of knowledge, when viewed by any individual from the
heart specialist's perspective, is termed explicit knowledge.
When the same notes are referred to, by a neurosurgeon for cerebro-vascular accident
diagnosis, the explicit knowledge created through these same notes becomes information
(since the current interpretation of the notes has lost its creation context). Therefore the
expressed knowledge viewed in the knowledge creator's context is explicit knowledge, but
not otherwise (Rodrigo 2001). This completes the knowledge cycle which can be
dichotomised into distinct sets of activities, one related to knowledge creation and the other
to knowledge sharing. Such knowledge cycles overlap and feed on each other in
propagating knowledge through the time continuum.
Knowledge creation
The dichotomisation of activities assisted in addressing the issues related to Knowledge
Creation (KC) and Knowledge Sharing (KS). Tacit knowledge creation within a human brain
can be viewed as a combination of input stimuli, even though there have been different
foresights in explaining how knowledge is created in cognitive environment from a
biological and psychological sciences point of view (Binnie and Williams 2002). This
discussion would limit itself to the cognitive side of knowledge creation as a black box and
rather focus on the inputs and outputs of human cognition (Schwenk 1988, Thomas et al.
2001).. In general the inputs to knowledge creation can be viewed as a set of stimuli which
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
15
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
can be classified into four main groups- explicit knowledge, interaction, information, and
insight – which coalesce in a continuous spiral process (Figure 9) to create new knowledge
(Baskaran et al. 2004). The external stimuli received by the human brain through its various
sensory organs can be in the form of explicit knowledge where the knowledge consumer
views the knowledge from the creator’s context (Figure 8). This could range from
physician’s notes to a financial analysis report. The interaction stimuli are related to learning
through physical experience.
Insight
Information
KC
Explicit
Knowledge
Interaction
Figure 9 - Knowledge creation spiral
The physical experience can be an articulated exchange of knowledge with an expert or
could be gained through physical experience such as learning to drive a car through
heuristics. Information stimuli are related to knowledge creation through the 1st G or 2nd G
information (Figure 8). This stimulus depends on the knowledge receiver’s ability to
comprehend the information using (self) past experience. Insight originates within; hence it
could be viewed as an internal stimulus. This cognition-based stimulus can be attributed to
all previous information, knowledge and wisdom possessed by the individual. It also has a
genetic aspect of knowledge retention. The above said stimuli in combination or alone create
new knowledge (Baskaran et al. 2004).
Knowledge Management (KM)
The earlier sections described in detail the different flavours of knowledge and stimuli for its
creation in a cognitive environment. This section briefly describes the concept of KM. In
contemporary terms, the view of KM can be summed up as “The process of lowering the
transaction costs associated with creating, sharing, and applying knowledge, and
developing improved strategies to support these activities” (Prusak and Matson 2006)..
After careful evaluation it can be concluded that humanistic factors can be attributed to the
intricacies experienced in present day KM (Prusak 2001, Prusak and Matson 2006, Ichijo and
Nonaka 2007). The process of knowledge creation and sharing is inseparable from its human
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
16
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
cognition (Ichijo and Nonaka 2007, Pemberton et al. 2001, Baumard 1999, Tsoukas and
Vladmiron 2001). Failing to appreciate the importance of these factors has misinformed
organisations embarking on a KM strategy (Walsham 2001, Wernerfelt 1984, Swan et
al.1999).
KM has been construed and viewed in different perspectives, some view as synergy, some
as Intellectual Capital (IC), competitiveness, a decision making tool, customer focus strategy,
employee relation and development, lowering costs, customer value addition, etc
(Chourides et al. 2003, Sveiby 1998). Healthcare organisations have started to believe that
long-term profitability and business sustainability are primary concerns and are to be
addressed outside the individual healthcare domains. So to achieve this, an organisationwide KM strategy has to be adopted (Ichijo and Nonaka 2007, Kasvi et al. 2003).. This
suggests that KM has to be viewed more in terms of leveraging than managing. Limited,
objective-based exposure has removed knowledge from a holistic concept and forced
business managers to view it as a resource which can be managed (Wernerfelt 1984, Swan et
al. 1999). A more proactive approach would refer to the unbounded subjective nature of
knowledge and rather conclude that knowledge should be leveraged (Prusak, 2001). Since
KM has been in use for over a decade and has been well established in both academic and
organizational domains, it will be appropriate to term all knowledge activities under KM, in
spite of the fact that KM should focus on leveraging knowledge (Walsham 2001, McDermott
1999).
Knowledge creation is still considered entirely as a cognitive process (Thomas et al. 2001,
Prusak and Matson 2006, Bell 1999, Ichijo and Nonaka, 2007, Pemberton et al. 2001, Baumard
1999). This process has not been completely understood by modern day biological science
and hence forces knowledge workers to look at it as a black box. From an organisation’s
perspective, the focus should be on the conducive environment for knowledge sharing to
create new knowledge (a knowledge cycle) among individuals within the organisation
(Thomas et al. 2001, Baumard 1999). Hence importance has to be given to knowledge cycles,
thereby assisting in the transcending process of creating a knowledgeable organisation. KM
oriented activities, strategies and processes aim at achieving a favourable environment that
implicitly propagates productive knowledge cycles.
The objective nature of knowledge was seen as an opportunity by many KM solution
providers. In the 1990’s such KM solution providers exploited organisations which showed
interest to invest in their readymade technological based KM solutions (Prusak and Matson
2006, Walsham 2001, McDermott 1999). Irrespective of the huge investments incurred in
implementing such technology-based KM solutions they utterly failed to deliver the much
promised returns on such huge investments. This resulted in a negative view on KM;
progressively organisations were wary about committing to future KM initiatives (Walsham
2001, McDermott 1999). This caused overexposure on non-core KM issues, which deprived
KM of the much needed maturing time. This misunderstanding can only be corrected by
refocusing KM on its precise core issues i.e. humanistic aspect. Liebowitz et al. (2005)
acknowledges the importance of people and culture as 80% in successful KM and attributes
the remaining to technology. In spite of the problems there is a growing interest in KM,
particularly in healthcare organisations. They have started to realise the potential of KM by
concentrating on the humanistic attributes of knowledge creation and sharing (Davenport
1995).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
17
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
KM has been accepted as a strategic management science (Drucker 1993, Argyris 1992, Senge
1990). Especially in organisations where exigencies as mentioned in the previous section are
profuse, KM is looked upon as the antiphon (Disterer 2002). Organisations have recognised
both sticky and leaky features exhibited by knowledge (Brown and Duguid, 2001).. Earlier
sections enumerated the nature of knowledge and the incidence of it can only be manifested
through humanistic traits (Ichijo and Nonaka 2007, Walsham 2001). In addition to this, any
organisation, especially healthcare is identified as people centric. This underlines the
importance of human nature within knowledge in organisations (Walsham 2001, Chourides
et al.2003). Every individual’s physiognomy has its own complexion- rules, morality and
economic associations, and this renders a unique phenomenon attributed to that particular
individual. Any organisation can be viewed as a layered model from the knowledge
perspective (Figure 10) (Ichijo and Nonaka 2007)..
Figure 10 - Multiple facets of Knowledge model in Project-based organisation
The core of the model is human psychology, as mentioned earlier all knowledge originates
at this level. Knowledge creation also takes place at this level. Human psychology either acts
as a knowledge enabler or a deterrent. Human values and beliefs are important factors for
urging the individual to share knowledge withheld in the individual’s cognition. The
stickiness manifests in this layer, this can be attributed to the cognitive facet of knowledge
creation. The knowledge leakiness can also be associated to this layer as it is related to
individuals leaving the organisation. At this level, Maslow’s theory of human needs can be
an ideal mechanism to explain and draw parallels regarding the motivating factors for
emphasizing knowledge creation and sharing (Maslow 1943). There are different theories
that explain factors which have a strong relationship in explaining an individual’s view on
knowledge sharing with other individuals from a personal perspective (human psychology
layer- refer to Figure 10). Personal level ego based human attributes influence the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
18
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
individual’s intent to share knowledge (Durkheim 1958). These factors are not permanent
and may fluctuate based on the environment to which the individual is exposed and its
combinatorial effect on the individual’s cognitive framework.
The second layer in the model is social psychology (Figure 10); this plays a major role in
sharing knowledge and in creating knowledge at group level. Some knowledge theorists
concur that “the sum is greater than the individuals” (Bresnen et al. 2003). Any
organisational team is a permanent/semi-permanent or temporary agglomeration of
individuals with a common purpose of achieving organisational objectives in time and to
deliver the set goals within stipulated resources; such organisational structure creates a
social bond among the team members. Even though sociological aspects can only be
deduced from a psychological perspective, a collective sense can stem from human’s social
nature.
Epistemic barriers can arise due to the complexities existing within a formal organisation.
Polanyi argues that social component is mandatory in all forms of knowledge creation and
sharing (Polanyi 1966). This concept of belonging and inter-responsibility can have an
adverse impact on how as a group the team will share the acquired knowledge not only
created within the organisation’s temporal continuum but also knowledge already
possessed by the individuals. The individual’s cultural, ethnic and education background
affects the intensity with which the individual will share knowledge (Fernie et al. 2003,
Brown and Duguid 2001). The individual’s inclination and eagerness to share knowledge
can be dampened by the current organisational culture. It can aid or deter the individual by
creating social apprehensions, peer pressure and aloofness among the group (Fernie et al.
2003, Brown and Duguid 2001, Newell and Huang 2005).
The concept of “Communities of Practice” (COP) is mostly constituted among peers who are
equal in stature in the organisation; this does have its inherent advantages such as
uninhibited knowledge sharing through cross-fertilization of ideas (Brown and Duguid
2001). Individuals pool their knowledge resources and act as temporary think-tanks within
the organisation; this can be mapped to the social psychology layer (Figure 10). This type of
COPs can be sustained beyond individual activities for which these COPs were constituted
within the organisation; thereby creating a forum which can be looked up for finding
solutions which are difficult from an individual’s capacity (Newell and Huang 2005,
DeFillippi 2001). These groups would put the organisations betterment over individual’s
gain. Experts concur that lack of proper recognition and appreciation will abrogate the
“stickiness” of knowledge, but on the contrary, this will facilitate the generation of
knowledge-islands in organisational environs. Incessant support to such knowledge-islands
creation would ultimately result in ineffectiveness of the organisation’s functional group
due to a decline in innovation and creativeness. The stickiness can only be abated by
supportive strategies addressing the social and communal apprehensions rather than the
individual’s beliefs (Hall and Sapsed 2005). This discussion would take the concept of social
groups such as COPs to the next level by expanding them to play a bigger role in sharing
knowledge irrespective of the individual’s position in the organisation’s domain (Hall and
Sapsed 2005). All the organisational staff should be forthcoming in sharing their knowledge
at the right time which can make a huge difference in the organisation’s success or failure.
The knowledge sharing culture should be viewed as a holistic characteristic of the team and
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
19
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
the organisation has to cultivate such an environment which can thrive and sustain a cordial
atmosphere for expansive knowledge sharing.
The next layer in the model (Figure 10) refers to the organisation’s cultural environment
which is vital for creating an enabling atmosphere for knowledge creation and sharing
(Lytras and Pouloudi 2003, Stonehouse and Pemberton 1999). This atmosphere should cater
to the psychological stratum at both the personal and the social level (Fernie et al. 2003). A
strong and dedicated leadership and top management which has self belief in knowledge
creation and sharing as an intrinsic organisational activity are mandatory (Marr 2004). Only
dedicated management which sets itself as role model can propagate a knowledge enabling
culture. Moral support and encouragement originating from the top management alone can
sustain a knowledge environment throughout the organisation (Marr 2004).. Durkheim in
his book on social theories postulates that when human subjects are intentionally or
unintentionally made to follow a routine and repetitive enforcement of constraints this
would gradually give rise to habits and internal tendencies which render the constraints
unnecessary in future (Durkheim 1958). Similarly, routine knowledge creating and sharing
attitudes can be ingrained within the organisation’s work ethics. This would not only ensure
a holistic application of knowledge but also enable new individuals joining the organisation
to be overwhelmed by the positive culture towards knowledge and embrace it
wholeheartedly, which results in the inhibition of psychological aspects in the individual
and a social approach alone will be gleaned out (Durkheim 1958).
Through these features, the organisation can realise its potential related to the beliefs,
tendencies and practices of its members. This eventually culminates into a truly knowledgebased social phenomenon. The organisation’s infrastructure has to be viewed as an enabler
for the knowledge environment (Stonehouse and Pemberton 1999). Many simple tools such
as an e-notice board, e-fora, blogs, discussion boards, mailing system, and a database for
useful information are vital (Wernerfelt 1984, Swan et al. 1999, McDermott 1999, Liebowitz et
al. 2005). Providing access to facilities for knowledge retrieval and sharing are fundamental
at this level, but every organisation has to realise that technology can play only a supporting
role and cannot replace the humanistic core of KM (Wernerfelt 1984, Tissen et al. 2000,
Liebowitz et al. 2005).. If a healthcare organisation can focus on the aforementioned aspects
of a knowledge-based organisation model, it can harvest many positive outcomes. The
outcomes are not only limited to a better execution of activities within the organisation, but
also increase the values of the organisation’s different facets such as Intellectual Capital (IC),
Social Capital (SC) and Knowledge Capital (Tissen et al. 2000, Prusak 2001, Marr 2004).
Knowledge Capital is related to both forms of knowledge i.e. tacit and explicit, including the
1st and 2nd G information. These capitals are fundamental for a successful, modern
organisation (Edvinsson 1997, Fong et al. 2005). All the above mentioned strategies can work
in unison to find solutions for the exigencies mentioned earlier.
Beyond KM
Two conflicting perspectives were discussed by experts, Wernerfelt (1984) views knowledge
as a resource that could be managed as any other in an organization. Wernerfelt (1984)
further details that knowledge exists as an independent entity from its creator and creation
context. Where as, Berger and Luckmann (1967) view knowledge as a culmination of shared
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
20
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
belief, prospering in social domain and extensive social interaction enhances knowledge
creation (Cicmil et al. 2005). This discussion draws its strength from Berger and Luckmann’s
views and expands and details the importance of humanistic core of knowledge. SECI
framework (Socialization, Externalization, Combination and Internalization) of Nonaka and
Takeuchi (1995) is acclaimed for its focus on socialization as a tool for creating and sharing
knowledge (Ichijo and Nonaka 2007)..
Some of the challenges which are to be addressed in organisational environs are as follows
(Ichijo and Nonaka, 2007) :
•
•
•
•
•
•
Knowledge crossing boundaries between team members of different
disciplines
Invisible boundaries between hierarchical management layers
Inter-disciplinary knowledge sharing is stronger than intra-disciplinary
Group discussion can create new knowledge generation
Integration of knowledge from different stakeholders
Encouragement to collective learning and inter-organisational learning
When the various humanistic knowledge aspects mentioned earlier are addressed
appropriately, any organisation can evolve into a knowledge-based organisation. As
solutions, many initiatives such as rewards, recognition, employing knowledge gate
keepers/enablers have been suggested (Fong et al. 2005, Cicmil et al. 2005). Organisations
and academia had accepted that rewards and recognition would encourage knowledge
sharing, but this works only on a short-term basis (Fong et al., 2005). These schemes of
appreciation are difficult to measure and benchmark, since the knowledge being shared /
created is beyond quantifiable mechanisms due to its rich tacit knowledge content. This may
force the organisation’s staff to take advantage of knowledge sharing as a means for
personal benefits, rather than for the betterment of the organisation as a whole. This
discussion on factors beyond KM, highlights the fact that all knowledge managed staff
should come forward to indulge and propagate knowledge sharing eventually creating new
knowledge without expecting specific compensation for such acts. They should accept
knowledge sharing as a culture and discharge knowledge creation activities as a part of their
regular activities. Hence a holistic concept of knowledge sharing and acceptance are
imperative for all future knowledge managed organisations.
Some organisations had employed a knowledge worker as a gate keeper of knowledge and
they are expected to not only assist in knowledge creation but also enable the free flow of
knowledge across organisational boundaries (Walsham 2001, Bresnen et al. 2003, Marr,
2004). But this approach is only intended as a short term arrangement and ultimately all
organisations have to embrace knowledge and do sharing as a routine process. A futuristic
knowledge sharing can be envisaged as in Figure 11. After the extended use of knowledge
workers in expediting their knowledge-based activities they themselves can become
obstacles for the free flow of knowledge.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
21
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
HC Organisation
Care group 1
Care group 2
Knowledge
gate keepers
Hospital
management /
surgery / PCT
Restricted
Knowledge
paths
External
partners (labs,
IT services,
contractors
etc.)
Unrestricted
Knowledge
paths
Knowledge
sharing
Figure 11 – Organisational knowledge sharing (inter and intra modes)
In addition, they will become liabilities to a knowledge system (Bresnen et al. 2003). Their
own personal egos, moods and prejudice can hamper the free flow of knowledge (Lytras
and Pouloudi 2003). To overcome the above-mentioned shortcomings, knowledge workers
have to be replaced by the entire team (executing a knowledge worker’s function as a group)
and thereby the organisation can realise an unrestricted flow of knowledge paths not only at
the intra level but also at the inter-organisation level. This can easily be expanded into
organisation wide knowledge flow paths and even parent organisation’s conglomerates and
partners can involve themselves actively in the organisation’s free knowledge sharing
(Cicmil et al. 2005).. Healthcare Organisations should not only commence KM based
initiatives but should be focussing more on sustaining such initiatives for a longer
timeframe. This allows the whole organisation to address new challenges by creating new
knowledge and sharing them as part of their routine activities. That is when one can realise
a complete knowledge focussed healthcare organisation. Care has to be exercised in
addressing all the facets of the knowledge model and focus at all levels of the model’s layer
to maximise benefits (Figure 11). Many KM models and frameworks have been proposed by
experts specifically for healthcare-based organisations (Whelton et al. 2002, Prencipe et al.
2005). The current discussion highlights that such organisations have to select the best
framework for their environment; they may even employ a hybrid of more than one
framework for maximum efficiency (Whelton et al. 2002, Prencipe et al. 2005). Irrespective of
what model/framework is selected, due focus on humanistic aspects has to be dealt at the
core level for achieving the maximum efficacy of KM. The model shown in Figure 11 could
be used as a guide to orient such KM models on humanistic aspects.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
22
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Total Knowledge Management in healthcare (TKMh)
Current KM strategies in healthcare are myopically focussed on how to convert knowledge
from a tacit to an explicit state. The Total Knowledge Management in healthcare (TKMh)
model, although abstract in nature, emphasises that KM-based implementations in
healthcare should be oriented towards the key issue of sharing knowledge in a tacit-to-tacit
mode. TKMh highlights the necessity of a human-centric approach to be adapted while
implementing KM strategy in healthcare. The model can be best visualised as a series of
healthcare management phases (as depicted in Figure 12). The "Initiate" phase brings about a
change in the environment to cultivate healthcare personnel's abilities and create tacit
knowledge impulsively as a routine healthcare-related activity. The second stage, "share",
sets up the environment so that healthcare specialists have the proactive desire to share
knowledge and focus more on tacit-to-tacit knowledge transfers.
Tacit knowledge
created
Initiate
Knowledge sharing
culture flourishes
Share
Returns
Establish
Exploit
Figure 12 - TKMH framework
The "establish" stage signifies that a culture has to be established to generate a network
structure and to sustain an uninhibited sharing of knowledge throughout the organization.
This should not be limited to one community, function or problem domain. The "exploit"
phase completes the iterating loop and would start when the healthcare organization can
garner the returns of the established culture through better realisation of its resources in all
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
23
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
forms and deliver better service to the public. The set of phases generates a continuous
feedback loop to its earlier stages and re-strengthens the iterating cycle until each phase has
attained its highest efficiency. The author believes that any healthcare organization, while
adapting a predefined KM framework, should analyse the model's individual components.
This would assist in selecting only the necessary individual components thus aligning the
framework to suit that particular healthcare organization's unique way of functioning.
Mapping these components to the TKMh model would provide overall guidance for the KM
implementation. This would correctly orient the implementation towards a tacit-to-tacit
mode of knowledge sharing. The term TKMh should not be confused with TQM (Total
Quality Management) as it is unrelated to this legacy quality concept. TKMh is an abstract
model aimed at implementing Knowledge Management in healthcare organizations
(Baskaran et al. 2004).
Knowledge creation and application of TKMh model for this project
Earlier sections provided the necessary focus on KC and sharing. This section applies the
generic KM concepts to the breast screening domain. In doing so evaluates and validates the
proposed KM concepts. As mentioned in section, various stimuli are mandatory in order to
instigate KC. As KC is fundamentally a cognitive process manifested within the human
mind, a favourable environment is required to create new knowledge (Figure 13). This
aspect could be best handled only by an organizational strategy that has faith in efficient KM
procedures. The change program currently taking place within the NHS is rapidly moving
towards this ideal (Baskaran et al.2004).
Figure 13 - Knowledge Creation in predicting screening attendance
The NHS Connecting for Health program places emphasis not only on creating the necessary
knowledge, but also provides an infrastructure to effortlessly share knowledge within the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
24
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
NHS’s various clinical pathways (Baskaran et al. 2005). The current research is aimed at
automating the process of prediction (the KC process) and efficiently sharing predicted
knowledge with primary care deliverers (GPs). Figure 14 depicts the application of TKMh
framework for the current research. In the ‘Initiate’ phase, the tacit knowledge creation is
commenced by an expert researcher. Research Methodology highlighted the importance of
qualitative research and, in this phase, knowledge from the previous researcher was collated
by way of semi-structured interviews and open-ended questions. This knowledge is later
captured by the AI algorithm for enabling continuous knowledge creation (prediction).
Initiate
Tacit knowledge creationfirst by experts and later
captured by AI algorithm
Share
Share tacit knowledge –
by HL7 messages
through NCRS
architecture
Humanistic value recognitionprimary care (GPs) aligned to
interventions for increasing
screening attendance
Establish
Exploit
Figure 14 - TKMh frame work applied for current research
During this phase of knowledge capture, the efficiency of the algorithm was
compared with the previous AI algorithm through triangulation. As part of the
quantitative analysis, features related to the receiver operator characteristics were
also completed. In the ‘Share’ phase, the created knowledge is shared by utilising the NCRS
infrastructure through HL7 messages bundles. The ‘Establish’ phase, recognises, creates, and
sustains the humanistic values related to the research domain. This phase focuses the
primary care deliverers (GPs) to make appropriate interventions. During this phase, the
qualitative research aspects were addressed through piloting within which openended questions and semi structured interviews were conducted with randomly
selected GPs for establishing the nuances of knowledge sharing. This assisted in
fine-tuning the proposed knowledge sharing mechanism and helped to ascertain the
challenges that should be addressed before this protocol can be fully utilised.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
25
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The data collected during this qualitative research stage warranted the need for
additional and comprehensive quantitative research (through a questionnaire-based
study) to gather as much information for leveraging the proposed KM-based
techniques. This would not only accumulate evidence but would also consolidate
the healthcare process-related changes necessary for the proposed protocol. This
provided an opportunity to triangulate the data collected during the piloting phase
interviews with the questionnaire survey findings. The ‘Exploit’, phase would signify
the continued efforts by not only primary care but also through other care deliverers to
address screening attendance. In addition, the infrastructure can be easily extended to cover
not only breast screening attendance but also other national screening programmes. An
iterative feedback loop assisted in reinforcing the project ideologies and fine-tuned the
research activities.
Knowledge Creation for this research is illustrated in Figure 13. The KC process is
instigated by the various forms of stimulus. These stimuli can be mapped to the earlier
concepts namely insight, information, interaction and explicit knowledge. The insight
stimulus comes from the initial idea of the researcher (the human component) to use
automated prediction (prediction analogy) as an effective tool to counter non-attendance.
Explicit knowledge acts as a proactive stimulus through knowledge derived from earlier
research by the group (on breast screening and AI) and through literature review of various
books, journals and academic materials. Information pertaining to the screening history of
the women and demographic details were collated from various sources. This constitutes the
information stimulus for this research. Interviews and meetings with the BSU’s staff and
previous researcher created the necessary interaction stimulus for KC.
The knowledge created by the various stimuli should not be a one time process (where
knowledge created would be static). This project work envisages the use of an AI-based
Neural Network (ANN) to capture the KC process. In this way, automation can be
implemented within the existing breast screening process; where the knowledge created is
in the form of screening attendance prediction. The ANN has been implemented using
Open Source technologies (Java environment). The implemented ANN works via a Javabased server program. It uses historical screening data and demographical information by
way of Townsend deprivation scores (Arochena 2003) as predicting factors (Adams 2004).
This forms the dataset that is presented to input neurons. The middle layer (also known as
the hidden layer) is connected to every other input neuron and to one output neuron. The
output neuron remains at “Zero” when a woman is predicted as an attendee and turns to
“One” when she is predicted as a non-attendee. Earlier research has confirmed that one
hidden layer will suffice to map any multivariate type of input domain to the output domain
(Arochena 2003).. During the training stage, the error function is fed back through the
network from the output neuron. The proposed knowledge capture is implemented using
the architecture shown in Figure 15. The input layer is presented with different
combinations of attributes based on the specific episode on which the network is being
trained. This simple back propagation based ANN is adequate for testing the above
knowledge capturing event..
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
26
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 15 - Screening attendance prediction through ANN
The AI-ATT prediction algorithm by Arochena (2003) channels the input through both the
pruning network and the RBFN. The final prediction outcome was based on better
prediction confidence between the two networks. The input dataset is passed through the
corresponding pre-trained network, based on the screening episode number. Extensive
network training and testing indicated that the more prediction parameters present (in the
form of historical episode details), the better the prediction results. The current research
relies on a similar ANN for prediction, it further recommends two approaches to share the
created knowledge with the primary care stakeholders (Figure 16). The proposed JAABS
algorithm also employs pruning network and RBFN.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
27
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
SOAP Message
Messaging server
(HL7 message
generator)
Breast Screening
Programme Server
(Runs JAABS &
NBSS)
BSP
Database
BSP Clients
N3 (formerly NHS NET)
GP Surgery Server
(Runs GP application)
Integrator
(Manages screening
appointment)
Other Clinical
Services with PCT
& Report
Generation
GP Clients
GP
Database
Figure 16 - KM-based architecture for increasing screening attendance
Knowledge sharing for this project
Even though KC through proper personification is essential, it would be of little use to an
organization unless it is properly mapped to all probable user-interactions within healthcare
services. The sharing process should also encourage the healthcare providers to carry out
their clinical and ethical responsibilities. Proper recognition (including monetary incentives
as short-term catalysts) could be used to encourage continuous improvement of
personifying knowledge at the delivery level. The knowledge gained through prediction
can best be utilized at its point of origin (breast screening unit). A priori knowledge of
predicted non-attendance could probably be used in breast screening unit to address empty
slots or to arrange standby screening slots or even to use the available equipment and staff’s
time suitably.
This knowledge sharing approach within the screening unit has limited scope for improving
its efficiency and on the other hand will have little or no progress on its resources front.
Additional resource gains can be made when this knowledge is shared in the most efficient
way with primary care providers, i.e. GP surgeries. This would provide a unique
opportunity for GPs to intervene and increase the screening unit’s uptake (Bankhead et al.
2001). The knowledge sharing process should also address the packaging and secure
transmission of knowledge to the surgeries. The GP surgeries may have a better chance of
making contact with women even before the screening date. This could instigate a GP
intervention. The intervention can be carried out by various methodologies such as:
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
28
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
1.
an arranged meeting with the women in person
2.
a telephone-based intervention by the GP
3.
using an opportunistic intervention when a woman approaches the surgery for
some other health-related service
4.
to send a letter signed by the woman’s GP explaining the importance of the
screening and emphasising the necessity of attending her allotted screening
appointment (this letter might be accompanied by more information about
breast screening and leaflets in the woman’s native language)
These interventions are aimed at educating woman on breast screening and would cultivate
awareness among probable non-attending women thus enabling them to make informed
decision. The quantitative studies accomplished through the questionnaire-based survey
shed more light on the preferred type of interventions and the suitable healthcare personnel
for initiating interventions. This project, as part of the personification process, addresses the
challenges encountered at the surgery and recommends possible strategies to make such
interventions an absolute success in increasing screening attendance. Only then can the
research can be deemed as a successful KM project. Based on this personification strategy
the proposed BSAMP protocol can be viewed in two facets. The first facet is related to KC
and the second facet is the knowledge sharing (attendance prediction is at the pre-screening
and screening results are at post-screening stage). Details of these facets are deliberated in
the following section.
Proposed twin approach
Earlier sections enumerated the current activities in a screening programme through a set of
‘use case diagrams’. The changes proposed and their relationships with existing activities
are shown in Figure 17. This figure also depicts the proposed sequence of actions involved
in the KC and knowledge sharing process within a breast screening centre. A marked
difference between these set of activities to the earlier ‘use cases’ described in earlier section
is the absence of PNL and its related processes. PNL has been discontinued after careful
evaluation and research by the National Screening Programmes (Codd, 2006).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
29
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 17 - Increasing screening attendance-twin approach
This section explains the BSAMP protocol for the KC (phase 1) and KS (phase 2) (Figure 17).
Data pre-processing refines the input dataset for ANN utilisation. When the prediction
algorithm is presented with the refined dataset, it predicts the non-attendees list. These
activities are identified in Figure 17 as phase 1.
Prediction and post-screening results
The first approach (Figure 17) refers to the following main actions: the Breast Screening Unit
(BSU) manager generates the incoming screening batch. This batching activity is executed on
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
30
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
a three year recursive cycle in such a way that it includes the whole eligible female
population covered by the screening unit.
When the batch list is generated by the NBSS software, the manager utilises a report
template to export the batch list as a flat file and saves it in a suitable folder located in the
BSU server. The next action is to generate the non-attendance list through the automated
software. The software component is resident in the BSU server and is accessed by the client
program via a Graphical User Interface (GUI). The software prompts the user to point to the
location of the flat file (already created by the template and stored in the server hard drive).
Once the file is located, the pre-processing software will automatically generate the required
predicting factors and normalise the data. The User (via the GUI interface) points the ANN
to the location of the historical data (in the flat file) to train the network. Once the training is
completed, the net is pointed to the test data so that prediction can be initiated.
Any errors during the pre-processing, training and the actual prediction activities are stored
in individual log files which can be viewed at a later point in time for feedback. The GUI
gives the option to the user to initiate the Simple Object Access Protocol (SOAP) message.
The message body is instantiated with reference to an eXtensible Markup Language (XML)
schema definition designed on the Health Level 7 (HL7) version 3.0 standards. The XML
message is called upon by the software to generate the SOAP envelope and attaches the
XML message to the SOAP body with a digital signature (for security). The Java-based web
services technology provides encryption to make the message completely secure. The
message is transmitted via web services to the GPs’ mailbox server (Figure 16). Once the GP
server connects to the mailbox it downloads the messages and the GP’s software
automatically updates the XML content to the women’s records after proper decryption and
verification of the digital signature contained in the SOAP message. Meanwhile, the BSU
executes its routine process of inviting the women by dispatching an appointment letter
(with details of the screening date and time).
For predicted non-attendees, the knowledge of non-attendance generated by the ANN is
now available to the GP. When the woman meets the GP for other healthcare related
services, GPs can proactively initiate an opportunistic type of intervention. Such
interventions can clarify women’s beliefs, attitudes by educating them on the importance of
breast screening; thereby they will make an informed decision and attend the screening
appointment (Bekker et al. 1999). The second approach (Figure 17) starts after the actual
screening batch. The NBSS creates the results for the last batch and utilises a report template
to export the batch list. The user (via GUI) points to the location of the flat file again to
segregate the non-attendees. Once segregation is completed, a new SOAP message is
generated using the same procedure as before and transmits it to the respective GPs. This
again updates the woman’s medical record of real non-attendance, thereby providing
another opportunistic type of intervention to the GPs.
From Figure 17, Phase 1 with a two-pronged approach proved to be an ideal choice.
Approach 1 is related to the knowledge created through the AI-ATT prediction algorithm
that predicts the non-attendance in the pre-screening stage. This approach represents the
process of duplicating the knowledge structure personified through the assistance of neural
networks. Approach 2 was related to the knowledge from the post-screening results. Both
results would be seamlessly integrated with the primary care providers. If the knowledge
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
31
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
from these approaches were to be channelled through an established primary care delivery
mechanism (such as a GP surgery), would yield better returns (Richards et al. 2001). Phase 2
is related to the knowledge personification at the primary care delivery end.
HL7 Message design
Based on a comprehensive requirements analysis, a conceptual diagram specifying classes
and their association with multiplicity for the scenario was constructed as the first step of
the message design (Figure 18). Screening units are responsible to screen the eligible female
cohorts. Each woman has many screening episodes within the specified age limit.
Women aged more than 70 are automatically out of the recall cycle but can still avail
themselves of the screening services by self referral; hence the number of episodes for each
woman cannot be limited to a logical number through multiplicity. Each screening unit
caters to many GP surgeries and each surgery is serviced by one or more physicians (GPs).
Each screening episode has one screening report. This report is mailed to screened women
(including non-attendees) and a copy of the report is also sent to the GP surgery with which
the particular woman is registered. This class diagram was further modified to reflect the
HL7 V3 Reference Information Model (RIM) paradigm.
•
•
•
•
•
•
•
•
GP Surgery
Screening Unit
• Surgery reference
Unit reference number
number
Address
1 Caters to 1..*
• Address
Signatory name
1
• Surgery name
Signatory position
• Telephone
1
1
Screens
Registered
with
1..*
Screening Women
NHS number
Address
Name
Telephone
1
1
Serviced by
1..*
GP
• GP reference number
• GP name
1
Copy sent to
Copy sent to
has
1..*
Screening Report
Episode
• Episode number
• Episode date
• Epsiode type
has
1
1
• Report reference
number
• Screening type
• Screening outcome
• Screening opinion
1
Figure 18 - Class diagram for screening scenario
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
32
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Modelling methodology
The sample screening report (hard-copy) sent to GPs was the basis on which the message
was designed (Figure 20). As the messages developed for the current research specifically
addresses the issues pertaining to non-attendance (and being prototype in nature), the focus
is on the two proposed scenarios alone (Figure 19).
Figure 19 - Design of message (screening) with objects and relationships
This shortens the scope and avoids the creation of D-MIM (Domain- Message Information
Model) As R-MIM is concrete and message specific in nature, the design phase concentrates
on creating the R-MIM (Reference Message Information Model) alone for this scenario. An
R-MIM pertaining to the pre-screening (prediction) and another related to the post-screening
is designed in this phase. The messages developed based on these R-MIMs will be validated
and tested through instantiation at the final stage of prototype implementation.
Information analysis
The twin message exchanges planned for this scenario have been developed after extensive
discussions and interviews with the participating screening unit staff and the paper-based
screening report sent to the GPs has been utilised as the main reference. Based on the
details depicted in Figure 19, two “storyboard” cases were created for each of the
knowledge transfers (refer to Error! Reference source not found. for more details).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
33
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 20 - Sample screening report sent to GP in hard-copy format
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
34
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The interaction diagram (Figure 21) refers to these storyboard cases for identifying the
communication between sender and receiver roles with related triggers. The naming of all
messaging components (roles, participants, triggers, interactions and messages etc) follows
HL7 naming conventions.
Breast Screening Office
PRSC_AR070001
General Physician (GP)
PRSC_AR070002
Run prediction
algorithm
PRSC_TE070001
Prediction Completed
Complete
screening
PRSC_TE070002
Screening Completed
Figure 21 - Interaction diagram for the twin messages
Information modelling
HL7 V3 message design is model-driven and is achieved through extensive use of graphic
models and each model represents a concept in message exchange. In concurrence with this
approach and, based on the storyboard, interaction diagram and tables corresponding to
RIM classes are cloned. The cloning of classes should reflect the semantic content of the
message intended for exchange. This cloning yields the first information model. After
meticulous iterations and incremental changes made to the basic model a final diagram
representing the complete semantic content of the message exchange can be drawn (Figure
19).. The objects reflect the colour scheme as recommended by HL7. The pre-screening
message (PRSC_MT060001 – refer to Table 2) is triggered by the trigger event
(PRSC_TE070001). The generic class groups related to the basic RIM classes are represented
as “swim lanes” for easy understanding and the lines connecting the objects show the
relationships between them.
Sending Role
Receiving Role
Trigger Event
Transmission Wrapper
Control Act Wrapper
Message Type
Breast Screening Office
General Physician (GP)
PredictionCompleted
Send Message Payload
Control Act with author
participation
Pre-Screening Report
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
PRSC_AR070001
PRSC_AR070002
PRSC_TE070001
MCCI_RM070001
MCAI_RM070001
PRSC_MT060001
35
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Table 2 - Pre-Screening interaction table (PRSC_IN060001)
The design process identifies the primary act as the “Screening (GP report)”. The roles
played within the act are:
•
•
•
the screening authority (sending role – PRSC_AR070001 – refer to Table 3)
sending the message
the GP as the message recipient (receiving role PRSC_AR070002)
service delivery location role identifies the place of screening (screening unit
or mobile unit)
Sending Role
Receiving Role
Trigger Event
Transmission Wrapper
Control Act Wrapper
Message Type
Breast Screening Office
General Physician (GP)
Screening Completed
Send Message Payload
Control
Act
with
participation
Post-Screening Report
PRSC_AR070001
PRSC_AR070002
PRSC_TE070001
MCCI_RM070001
author MCAI_RM070001
PRSC_MT060002
Table 3 - Post-screening interaction table (PRSC_IN060002)
The entities associated with the roles played are individuals represented as persons (A, B and
C) and the screening unit. The GP plays the message receiver role. An additional role is
assigned to the GP as the family physician for the screened women, which is represented
through role relationship “registered with” (Figure 19).. The act itself is made up of more
than one related act and this pertains to details such as screening type, screening opinion
and a note. These acts are related to the screening GP report by act relationship defined as
pertinent information for each associated act.
HL7 v3 Message implementation
The message designed in section 4.6 was implemented through the following steps using the
specified software tools. The initial message implementation was through UML diagrams
(refer to section). With reference to these diagrams, detailed RIM-based domain diagrams
were generated. The final design has been translated into an HL7 object model artefact using
MS® VISIO® (Professional 2002 v10.0.525) software using HL7 templates (RMIM2) which
loads the RIM and vocabulary for the attributes linked to these stencils (version 2.99). The
relevant attributes are selected from the RIM classes satisfying the mandatory conditions
stipulated by HL7 (such as Instance Identifier, class code, mode code etc) and those relevant
to the message content, with reference to the class diagram (Figure 18). Data types, coded
elements and multiplicity associated with the relationships were also defined during this
phase.
The HL7 UK and Connecting for Health (CfH) have well defined Common Message Element
Types (CMETs) within the domain. These CMETs aim at creating ready made blocks for
frequently reused components and messaging structures. This facilitates rapid message
designing and assists designers in not reinventing-the-wheel. CMETs also provide error-free
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
36
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
message design with consistency. To gain better understanding of the message designing
nuances and to gain more design experience (owing to the projects focussed context) the
author created some of the components exclusively for this work, in spite of the existing
equivalent CMETs. The wrappers necessary for the messages, Control act wrapper
(MCAI_RM070001 refer to Table 1) and Transmission wrapper (MCCI_RM070001 refer to
Error! Reference source not found.) were manually created with the VISIO® templates in
accordance to CfH MIM (v 3.1.04).
Drawing the control act wrapper as recommended by CfH is simple with the VISIO®
stencils, but the transmission wrapper similar to the recommended one has not been
possible due to the non-availability of certain library files. This created validation errors in
VISIO®; failing this validation, no Hierarchical Message Definitions (HMDs) were created.
Hence, the automated schema generator (HL7 V3 Schema Builder V1.20) is not able to
generate schemas. To overcome this problem, and to complete the message design for the
prototype, the wrapper schema has been hand crafted. The rest of the message components
are validated using the HL7 vocabulary and HMDs are built using “RoseTree” software
(version 2.9.55, RIM-V02-04). Schema Builder has been used to generate the XML schemas
automatically (refer to Error! Reference source not found. and Error! Reference source not
found. for a sample schema for the proposed messages).. Similar to the pre-screening
message and its components, the trigger event, message type and its associated sending and
receiving roles have been designed for post-screening interaction (PRSC_IN060002, refer to
Table 3). The same control act wrapper and transmission wrapper have been reused for this
message.
Challenges faced during message design and implementation phase
HL7 standard version 3 is a complex standard for novice users. Exposure to both healthcare
and software domains would greatly assist in understanding the more rudimentary aspects
of HL7 version 3. This project had proved that even a novice in EDI and HL7 could make
substantial inroads if the learning process were to be properly directed.
The following critical aspects were encountered during the research, this type of HL7 V3
message design process is recommended where the available resources are limited.
1.
Becoming a member in the local HL7 organisation has its own benefit. It is a
good place to meet and know the local thrust areas and the modalities for
local implementation. An important factor would be to get some form of
introduction to the standards through professionally designed training
sessions which are organised by the HL7 group.
2.
There is a sizeable volume of literature available at various sites, such as HL7
UK and HL7 (US, Canada and Australia to name a few).
3.
The HL7 parent organisation offers tools and technologies for developers and
message designers and this provides a conducive environment for beginners
to get a foot hold easily in design and implementation of HL7 messages.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
37
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
4.
Not only the HL7 V3 development tools are provided free for download even
CfH MIM (for members only) is also available for new message developers.
These support mechanisms also include necessary guidelines including
configuration and installation notes.
During the initial phase of developing the HL7 version 3, the Standards Development
Organisation (SDO) for HL7 had pursued an easier method for setting up the working
environment, hence had logically selected a well established, off-the-shelf tool such as MS®
VISIO® and added its own libraries through Visual Basic (VB) based macros. Rose Tree
offers the much needed interface to convert the VISIO® designs into HMDs and message
instances. These are exported as comma separated values (CSV) for further processing.
Schema Builder, a java based programme takes in these CSV files as input to generate XML
schemas automatically. There has been a migration towards the use of Eclipse as a
development environment and this has gained popularity but this is not mandatory for
message development.
Identifying the related HL7 library files and having access to them can be a daunting
challenge for any outsiders to the working group. Hence some form of collaboration with
the local (country level) sub-group is vital for any message development. This would avoid
a lot of hardship and align the projects with the local implementation. Even though the
VISIO® stencils, Rose Tree and Schema Builder are free to download, the need for MS®
VISIO® which is commercially available software is necessary for designing the messages
and this can be a handicap as it is not free to download. The literature and other related
papers are spread across different domains and thus require careful collation for effective
comprehension. The lack of exclusive academic courses on HL7 or healthcare informatics
has created a gap in training and educating fresh talents at graduate level in this field. In
spite of the freely available tools, there are a few barriers to be overcome as a beginner.
These tools were developed by individual groups within the HL7 SDO and hence function
as separate components. An enterprise-based solution which encompasses all the tools and a
single interface would be very useful. This would reduce the complexities of the message
design and implementation.
This particular message design itself was simple and closely corresponded to the paper based
communication between the screening unit and GPs. Mapping the contents and adding new
components to satisfy the HL7 syntax without affecting the semantics was relatively easy.
For novice designers, such message design can be a good starting point. As the designers
gain more experience, more complicated messages (for example, spanning multiple clinical
documents) within the domain can be designed with relative ease. The scope of this project
had a distinct focus and therefore provided a unique opportunity to complete the whole
messaging (design through implementation).. This research is directed at creating a prototype
level message only. The final prototype would be a complete software environment which
would incorporate a database input hooked to the AI network for predicting non-attendance.
The predicted women will be automatically associated with the respective GPs and the
message content would be transferred to the messaging module. This module will use the
message content and the schema (as designed and developed earlier) to generate XML
instances. These instances would be part of a Simple Object Access Protocol (SOAP)
message and sent to the respective GP system. This would later be utilised to initiate GP
intervention when the identified women approach the GPs for other health-related services.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
38
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
This section enumerated some of the challenges that can be encountered while creating HL7
messages in a highly focussed environment such as the prototype discussed. The author has
exhibited that HL7 development methodology could be learned and applied in a short time.
As the standards are maturing and more countries are adopting and implementing HL7 V3,
this research can be an ideal starting point for novice message designers and implementers.
Proposed protocol
Reasons for non-attendance may be largely attributed to disinterest in attending a
mammography session, negative attitudes, beliefs, prior or current medical problems, and
fear of X-rays (Bekker et al. 1999). These reasons can be negated by a proper education
provided to women. Education has to be directed at explaining the advantages and
importance of screening and assist in removing the socio-cultural and personal barriers
(Cassandra 2006). Other possible options including convenience in terms of time, place and
dates provided to women for encouraging their attendance. In spite of the expedient
measures provided to the women, non-attendance has been a grave concern for the National
Screening Programme. This scenario can be properly addressed if women can be identified
who may probably not attend in advance and directing additional resources to educate these
women, thereby increasing attendance. To summarise the messaging and the practicality of
the proposed solution, a detailed architecture utilising the existing technologies and
proposed components are recapitulated in this section as a new breast screening protocol.
The same architecture can be mapped for other message paths in the future.
Proposed Breast Screening Attendance Messaging Protocol (BSAMP)
The Health Level Seven (HL7) Standard is the de facto standard for all electronic exchange of
data with in healthcare environments in the NHS (Thorp 2005). The HL7 standard defines
seamless exchange of actual data (and its structure) between health care applications (Beeler
1998). This section addresses the BSAMP and its messages pertaining to a prototype model
created with the assistance of an Artificial Intelligence (AI) program (AI layer - refer to
Figure 22) to predict breast screening non-attendees and to share this knowledge with the
primary care deliverers (namely the GPs) and they are:
•
•
Pre-screening (predicted non-attendance)
Post-screening (actual non-attendees)
These two approaches were described in the earlier sections with clarity. The driving
component for both the approaches is the infrastructure required for the knowledge sharing.
The existing technologies described have to be utilised to reflect the IT revolution occurring
under the NHS’ national initiative. This would not only provide the path for faster
implementation but also reduce the required overheads by utilising the in-place system for
knowledge exchange. The proposed BSAMP protocol leverages these technologies as shown
in Figure 22. This illustration also points at five different message flow paths associated
with the said screening domain.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
39
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
2
3
4
5
1
Figure 22 - BSAMP Protocol architecture
‘Path 1’ is the existing messaging channel used by the current screening centre’s exchange of
screening report (at post-screening stage) to the GPs in hard-copy format. This path is
cumbersome not only to produce at the screening centre’s administration office but also
would require considerable manpower and resources at the GP end for entering the
information in to the women’s clinical records (either digital or paper-based).. The
remaining paths are all proposed to utilise the new IT infrastructure for other message
exchanges in concurrence with the suggested protocol. The screening centre collates all its
screening statistics in to consolidated information package for ‘Körner Community’ returns
(KC62 and KC63 reports). These reports are sent to the National Screening Programmes for
generating national statistics for all screening programmes. This information exchange can
be made using the same architecture as portrayed in Figure 22, ‘path 2’.
‘Path 3’ proposes to download and upload any changes to the women’s demography and
personnel information. ‘Path 4’ is intended to record the outcomes of the screening in to the
Personnel Spine Information Service (PSIS). This updating will provide an opportunity for
all the healthcare providers to gain knowledge of the screening outcome (including
attendance and non-attendance) so that a collective effort can be made to deliver better
health care. ‘Path 5’ is for sharing the knowledge of non-attendance gathered from prescreening (prediction) and post –screening report. The NBSS system communicates the
dataset to the AI layer for prediction, and the outcome is appropriately packaged as HL7based SOAP messages by the wrapper layer and then via the Data Transfer Service (DTS)
client interface, the message is sent to the DTS layer. This populates the respective GP mail
boxes.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
40
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Typically, on regular intervals (3 to 5 times per day) the GP application layer executes a
batch process of uploading and downloading the GP mail box through the DTS client
interface for GPs. When the screening report message arrives at the GP application layer,
after due authentication, the individual women’s clinical records are updated. There are two
main families of GP software for assisting GPs in their clinical consultations. One type of GP
software uses pop-up windows and another uses framed structures to draw attention
through textual prompts (Figure 23 and Figure 24)..
Figure 23 - Screen shot of prediction and non-attendance displayed in frames with different set of font
colours
When the woman approaches the GP for consultation, the GP retrieves the women’s record
and these trigger the automated prompts (Figure 24). The prompts are amber coloured and
pop up to remind the GP to intervene, since this woman has been predicted as a nonattendee. The other screen shot (Figure 24) depicts the red coloured pop-up prompt
signifying the non-attendance at post-screening stage. This incidentally adds value to the
already existing prompt (amber coloured) of predicted non-attendance. In the other type of
frame-based GUI, appropriate coloured text is used to highlight the importance of the
shared knowledge.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
41
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 24 – Screen shots of predicted non-attendance displayed as ‘amber’ and non-attendance as a
‘red’ pop-up prompt on the women’s clinical record
Conclusion
The BSAMP architecture relies on existing technologies. This gains significance as it implies
that this is an implementable solution. The results collected through the qualitative and
quantitative analysis served to further understanding as to how knowledge creation
and sharing can be established with reasonable efficiency. The AI triangulation
provided the efficiency metrics (benchmarks) for future improvements. The NCRS is
already being tested across the country. The architecture calls for close cooperation with the
screening units and other care deliverers. Such cooperative efforts can make significant
impact on screening attendance. This chapter describes how KM-based tools and techniques
can assist in evolving new healthcare strategies to address new challenges.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
42
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Algorithm for Screening Attendance
Prediction
The first step towards addressing the screening attendance was largely completed by
Arochena’s (2003) research on predicting non-attending women using an Artificial
Intelligence (AI) based predicting algorithm. The following section provides a summary of
the Arochena’s work (AI-ATT), a brief analysis of its functioning and working environment,
later the chapter focuses on the proposed (JAABS) algorithm’s data pre-processing modules
and finally dwells on the JAABS algorithm’s design aspects. As discussed earlier,
quantitative analysis was completed on the AI’s performance metrics. Further, these
results were triangulated to reflect how the new algorithm compares with the earlier
one and also to the statistical technique (logistic regression). This comparison of the
JAABS vs. AI-ATT is enumerated and concluded.
Predecessor algorithm employed in exploratory research
The research by Arochena (2003), funded by and tested in the Warwickshire, Solihull and
Coventry Breast Screening Unit, (project ID 2000/112) indicated that women’s attendance can
be predicted by a software system using the Artificial Intelligence based attendance predicting
algorithm (AI-ATT). The algorithm has been evaluated with individual AI methods (feed
forward and radial basis function), and statistical methods (such as Logistic Regression - LR)
and was found to be extremely favourable in predicting attendance (Arochena 2003).. This
algorithm can assist the Breast Screening Unit (BSU) to effectively schedule the screening
programme and eventually result in the efficient use of limited available resources. Even if
the attendance percentage were increased marginally, a significant number of lives may be
saved. Arochena’s (2003) research resulted in the formulation of the required artificial
intelligence-based attendance prediction algorithm and was trained and tested on datasets
collected during a ten-year period at the Warwickshire, Solihull and Coventry Breast Screening
Unit. The blind study (data not seen earlier by the AI) test results were encouraging,
especially from the second episode onwards, warranting the algorithm to possess enough
potential to be incorporated in a Breast Screening Programme (BSP) framework.
Wilson’s (1999) work focussed in developing a methodology to extract data from the
MUMPS database. The same methodology was later used to extract dataset pertaining to the
period of 1989 to 2001 from the Warwickshire, Solihull and Coventry Breast Screening Unit for
testing and validating the AI-ATT. In addition, the National Screening Programme has been
constantly striving to provide better services to the public and one of the new enhancements
offered by the screening services is to increase the screening age limit from sixty four to
seventy (Patnick 2006). This effectively increased the number of screening episodes and
resulted in augmenting the need for effective use of the already stretched NHS resources.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
43
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
All the above said factors raise the important question pertaining to how effectively the
predicted knowledge can be leveraged and managed.
An exploratory research exercise identified two algorithms, one related to predicting
attendance on screening batch and the other to predict the incidence of screening variation,
i.e. changes to the appointment dates, resulting in late attendance (Arochena et al. 2001).
The research further validates the algorithm through a blind study by using the latest data
from the screening unit (not formerly seen by the algorithm while training and testing)
(Arochena et al. 2003, Aochena et al. 2002). The results confirmed that these algorithms have
significant potential and were predicting with reasonable accuracy. This warranted further
research to confirm the feasibility of implementing such algorithms in the routine HC
delivery process at the screening unit.
Artificial Intelligence based Attendance Prediction algorithm (AI-ATT)
The earlier section gave a brief overview on the research completed by Arochena (2003). This
work had modelled the predicting factors and was validated through an Artificial
Intelligence (AI) based predicting algorithm. The dataset (containing 281,415 potential
screenings related to 147,432 women) was downloaded from the screening unit’s MUMPS
(Massachusetts General Hospital Utility Multi-Programming System) database (Wilson 1999,
Arochena et al. 2002). This data was filtered by a combination of Visual Basic® (VB) and
Sequential Query Language (SQL) based macros and routines in a Microsoft Excel
environment respectively. The routines used in this data filtering process were primarily
designed to identify unknown date of birth, irregularities in date entry, incomplete
screening end code, inconsistent episode numbers, anomalous SX number, and other
missing data, incorrect data entry, duplicate entries etc (Wilson 1999)..
Analysis of AI-ATT
Considerable association was acknowledged between age bands, attendance and screening
end code (Arochena et al. 2003, Arochena et al. 2003). Age band including round length
(difference in years between the date of first offered appointment of one screening episode
and the following one), previous screening attendance and screening variation had high
association.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
44
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 25 – Data mining stream for AI-ATT algorithm (Adapted from Arochena’s project report
(Arochena 2003)
Postal area, Townsend deprivation, screening variation, past cancer history, false positive,
history of false positive and type of invitation (first call or recall) to the episode also had
substantial relationship with the screening attendance (Arochena 2003, Arochena et al.
2003).. The predicting factors were not the same for all the episodes due to various reasons.
This prompted the input data to be split into their respective episode groups. The algorithm
has been evaluated and compared with Logistic Regression (LR) model and was found to be
more precise in predicting women attendance (Arochena 2003, Arochena et al. 2001). The
algorithm employs two approaches one through a neural network pruning and the other by
radial basis function based network. A hybrid of these two approaches was employed in the
AI-ATT to attain better performance metrics like accuracy, sensitivity, specificity and
positive & negative predicted values when compared to LR (Arochena 2003, Arochena et al.
2001)..
AI-ATT working environment
The analyses of relationships between the input variables and the attendance pointed that
not all the episodes had the same set of predictor variables. For example the first episode
which has all the relevant demographic details of the women lacked screening history which
was a handicap and yielded less accurate results (Arochena 2003). After exhaustive
analyses, it was found that the type of invitation and the age of the women were positive
prediction attributes for the first episode group (Arochena 2003).. Whereas the remaining
episodes comprised of more details (demographic and screening history) this enabled better
attendance prediction. The attribute indicating non-attendance in previous episode and
diagnoses of breast cancer in the previous screening episode were also identified as
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
45
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
predicting factors (Arochena et al. 2002). The algorithm functions are enumerated as steps
one through seven (Figure 26). The dataset as shown is derived from two different sources,
the local Primary Care Trust (PCT) holds the updated information of the women’s
demographic details and the screening unit holds the details pertaining to the women’s past
screening episode details. The screening unit downloads the demographic data from the
PCT to create the batch list for the upcoming episode. The algorithm then combines this
data with the screening history for available women. The data from this combination is
extracted as a single file which is in Microsoft compatible format for further processing.
The data is manipulated and reorganised based on the episode number and each woman is
identified with an episode number which is inserted as a variable.
Figure 26 - AI-ATT algorithm working process (Arochena 2003)
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
46
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Based on this inserted episode number the single monolithic file is partitioned into
respective episode groups and saved as sub-files (Arochena 2003). The postcode of the
women’s residence (extracted from the women’s demographic data) at the screening time
period is matched with the “Townsends” deprivation score and is stored as a new variable.
Townsend’s deprivation score is calculated utilising the census data. The time difference
between the census data and the physical expansion rate of a city results in areas that do not
have corresponding deprivation scores. It was found that not all the postcodes had an
associated Townsend’s deprivation score; hence based on the presence of this variable the
sub-file is further classified (Arochena et al. 2002).. A number derived from the postal area
enumeration based on district list is assigned for each record. This number is leveraged for
postcodes that do not have related Townsend’s deprivation score and is stored as a new
variable called post annum. The above said two sub-files (one related to Townsends
deprivation score and another related to postcode) are individually fed to two Artificial
Neural Network models (ANN) for every episode. ANN based on Pruning (ANNP) and
Radial Basis Function Network (RBFN) are the two ANNs employed. These ANNs generate
two individual variables; one related to the attendance prediction and another related to the
confidence value of the prediction (Arochena 2003).
The final activity is related to the classification of the outcome from the previous step. The
confidence level of the predicted attendance is compared between the values generated by
the ANNP and RBFN algorithm. The highest confidence level between the two is voted as
the final prediction of the AI-ATT algorithm. The final dataset is again collated from these
sub-files into one consolidated output file with the predicted attendance (Arochena 2003,
Arochena et al. 2000, 2001, 2002, 2003).. The dataset collected during the ten-year period had
many versions or styles which need to be regularised towards a common platform before
being put to use within the algorithm (Arochena et al. 2000). This data processing was
carried out using several different software data analysis tools. These tools included
Microsoft Excel®, Visual Basic® routines and Statistical Package for the Social Sciences
(SPSS®) and Clementine® to extract the data and adapt to the AI models. The algorithm
itself was modelled in a visual modelling environment called Clementine® (Arochena 2003).
The algorithm for its proper functioning depends on all the aforementioned software. This
posed a challenge in implementing the algorithm within the routine process of the breast
screening office environment. The primary challenge was that they do not possess the
required licenses for the said software nor the expertise in linking all the software to work
seamlessly for predicting on a monthly basis for every screening batch. The earlier research
by Arochena (2003) was limited to identifying whether the non-attendees can be predicted
with reliable accuracy and was not aimed at the various facets of implementing the
algorithm nor on envisaging how to effectively use the gained knowledge of probable nonattendance. This project work addresses such issues from an implementation point view
and suggests how this knowledge can be leveraged. The qualitative research conducted at
the initial phase through interviews fostered better understanding of the previous (AI-ATT)
algorithm’s developmental challenges and how they were resolved. This played a
significant role in reducing the development time for the new algorithm. The next section
enumerates the design and implementation of the new algorithm.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
47
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Data pre-processing
The introductory section gave a brief insight to the research by Arochena (2003), who had
pursued an exploratory type of research focused on identifying the attributes that can be
classified as predicting factors. The primary objective of the said research was to propose
the predicting algorithm which includes the modelling of the prediction features that affect
screening attendance to screening invitations. The current work elaborates the need for an
open source-based data pre-processing module (Figure 27). It further illustrates the design
of this module and the variables participating in the prediction process.
Figure 27 - Pre-processing module for proposed prototype
Data Pre-processing design
The demographic details for the three year call/recall are downloaded from the local health
care authority’s database. The downloading is effected via the health link network on to a
standalone system within the Breast Screening Unit (BSU). The historical data related to
screening, appointments and results pertaining to screening women are retained within the
screening unit’s “MUMPS” database (Figure 27). “MUMPS” also called the Oxford system,
is one of the earliest programming languages used since 1960’s (ÒKane 2002). This language
was extensively employed to write database applications explicitly for the HC domain.
The MUMPS database works on Disk Operating System (DOS) and employs a characterbased user interface for database interrogation (ÒKane 2002). The cumbersome DOS based
system was prone to erroneous data entry and hence warranted a change in the system.
New software, National Breast Screening Computer System (NBSS), was developed in 20023 to address these issues (Baskaran at al. 2005, Tarver et al. 2004).
This NBSS consists of a Visual Basic® (VB) front end connected to a “Caché” ® database
which was seamlessly integrated with the MUMPS database (Tarver et al. 2004). All the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
48
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
above processes required multiple software interactions, functioning on different
environments which resulted in considerable complexities during data extraction for the
current project.
Module design
The earlier research conducted within the BIOCORE had used the MUMPS database to
extract data. This data was saved as a flat file and then imported into MS EXCEL® for
further processing. The data was filtered by VB® application routines and SQL queries
which were manually executed and verified for its correctness (Arochena 2003, Wilson
1999). The filtering addressed Y2K (year 2000) bug, unknown date of birth, irregular data
entries, incomplete screening end code, anomalous episode number etc (Arochena 2003).
The above said process was tedious and human centric. Each and every record has to be
validated manually and incomplete dataset were excluded from the study. In the current
study very similar difficulties were encountered. In addition to these challenges, the
prototype proposed has to be completely automated. The automation has to address the
removal of anomalous data, wrong data entry, duplicate entries etc. All the experts’
involvement in data validation and analyses has to be expedited through the prototype with
little or no human intervention.
In an ideal scenario, the functioning of the proposed prototype (Figure 27) for attendance
prediction would encompass all the above data processing activities plus the generation of
the required predictor variables using the logic proposed in the algorithm. The data preprocessing module after completing the above said activities has to generate a new dataset
that would be ready to be assimilated by the AI module. The earlier research’s data
processing was conducted in different environments hence was difficult to be integrated as
one whole functioning unit. All the above said challenges can be solved if there is one
software module which can deliver a complete and automated data pre-processing for the
prototype. Hence an automated data pre-processing is indispensable for the successful
functioning of the proposed algorithm. All the activities pertaining to the proposed
algorithm can be summarised as one ‘use case’ diagram (Figure 28).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
49
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 28 - High level Use Case diagram of the prototype
Variables identified for data pre-processing
The data pre-processing module can be visualised as two separate sub-components. The
screening office module is executed with the existing software programmes available in the
breast screening office and the new Java-based pre-processor module for collating the data.
The VB® front end made data extraction straightforward (Figure 27) from the MUMPS
database through SQL queries directed at the Caché® database. Currently the breast
screening office is employing “Crystal Report” (CR) as part of the NBSS to generate reports
for all the screening activities including screening, administration, invitation etc. Part of the
data pre-processing was implemented through CR software. The screening unit had earlier
indicated that the routine functioning of the screening office should not be affected during
the data extraction process. Hence, prior to the data extraction a CR template was created to
reflect the format of the data to be exported. This template was used to export the data as a
flat file. This strategy was adopted since all the screening units across the country were
expected to have some form of minimum facility for creating dataset in a flat file format.
Coupled with this a need for a low overhead on the existing IT system and minimum
additional complexities were expected as fundamental for the proposed prototype. All the
aforementioned rationale strengthened the need for adopting a compromise strategy which
would export data as a flat file, so that the mode of data transfer can be standardised across
the country with minimum or no interrogation with the screening database. The CR
template generates the dataset as depicted in the format (Table 4). The query generated the
demographic and episode details for all the women in as many records. The demographic
data was not complete and only the first record of a particular woman had the complete
dataset and the remaining records of the women corresponded to the historical episode
details. The SX number was repeated one for the demographic data and one for each
episode.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
50
Episode end
code
Reason episode
closed
Clinic code
Screening date
Call type
Date of First
offered
appointment
Episode ID
SX Number
GP postcode
GP Surname
Postcode
Date of Birth
SX number
NHS Number
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Table 4 – Crystal Report Template for data extraction
The SX number is the link between the two components of the record. The women’s address
and name are excluded from the study to address data protection and to maintain
anonymity. In spite of the necessity of the messaging module the complete data set was
generated without the personal information of the screening women. The post code of the
women is indispensable for the current study, as it generates the important predictor
variable in the form of Townsend’s reference number and post annum number. The
recipient of the proposed knowledge transfer is the woman’s GP, hence GP information like
surname, surgery address and postcode are necessary for sending the HL7 based EDI
message. This dataset including the women’s GP details are collated as a separate entity.
For further use, it is stored as a look-up file for generating messages. The SX number acts as
the link between these two datasets. The women’s demography is represented as a ‘Record’
object (Figure 29). The screening detail for each episode is represented as ‘Episode’ object.
IO module
Figure 29 – Data Pre-processor module class diagram
One ‘Record’ object would be associated with one or more ‘Episode’ objects. The gaps in the
demographic record have to be filled and the episode details are associated with the
women’s demographic data. Exhaustive analyses of the data indicated that the CR report
had duplicate episode details. Upon further analysis it was found that rescheduling the
appointment by the women was the primary reason for such multiple and duplicate entries.
These entries are to be removed before further processing can be implemented. Each record
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
51
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
read from the CR report has to be first partitioned in to episode details and are stored as
‘Episode’ objects. They are finally collated and associated with the women’s demographic
details (represented as ‘Record’ object). In addition to this all the records has to be
automatically validated. The earlier work by Arochena (2003) had identified all the
contributing predictor attributes through comprehensive statistical analyses. Based on these
analyses, the following attributes are generated in this module.
•
•
•
•
•
•
•
•
•
Townsend reference number
Post annum number
Age band (age of the women for the current episode, converted as categorical
variable)
Previous screening attendance
Slip (difference between the date of first offered appointment and the actual
screening)
Number of tests (diagnoses corresponding to the immediately prior episode)
Cancer detection (in the immediately prior episode)
False positive (false positive diagnoses results in the immediately prior
episode)
History of false positive (false positive occurrence in any of the previous
episodes)
History of cancer (cancer detection in any of the previous episodes)
•
•
After generating these attributes the pre-processor module classifies the ‘Record’ objects
based on the number of ‘Episode’ objects it contains. This dataset is then written as an inprocess flat file for reference. All errors generated during the execution of the preprocessing module are written in a log (error) and is also saved as a flat file for future
reference.
Data pre-processing module implementation
The data pre-processing module is built with a sub-module called “IO module” (Figure 27)
for reusability. This module has three classes, one to create a file for a given name, another
to create a stream to read the file line-by-line and the last being the class to invoke a file
writing stream. The main class of the data pre-processing module is the “DataPreProcessor”
class. This class drives the whole module; a ‘Record’ class stores the attributes of the
screening women’s demographic details. The ‘Episode’ class holds the attributes of the
woman’s screening history details. The ‘Record’ class stores all the episodes it is related to,
in a “Vector” object. The ‘Record’ class uses “DateCompare” class for comparing dates to
store the episodes sequentially and to identify ‘Episode’ object duplicates. The same class is
used to generate the “Age band” predictor variable from the woman’s age and the date of
first offered appointment and “Slip” variable is also calculated using the same class by
finding the difference of date between the appointment date and the actual screening date in
days.
The “DataPreProcessor” class populates vector objects with the Townsend reference and
“Postannum” number by reading values form a flat file utilising the “IO Module”. The
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
52
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
postcode’s first part is extracted and used to check the availability of Townsend number
stored as Vector object. If the Townsend number is not available then an arbitrary number
(say 100) is stored as Townsend reference number. Similar to this a “Postannum” number
derived from the postal area enumeration is matched with the district list (from a lookup
file) and stored in the ‘Record’ object. The “DataPreProcessor” object first loads the input
flat file; the flat file could be either as coma separated values (CSV) or tab delimited values.
To load this file the class utilises the IO module’s “ProcessFileReader” class and reads line
by line and process one line at a time to create the ‘Record’ and ‘Episode’ objects
respectively until the end of file. After creating these objects, the predictor attributes are also
generated based on the input data and stores in their respective ‘Episode’ objects. The
“DataPreProcessor” object then saves the data by opening a file writer stream through the
“ProcessFileWriter” object. This file becomes the input to the subsequent AI module for
normalisation of data before feeding it to the neural network.
Proposed algorithm
During the initial stages, qualitative analysis revealed that the software platform was
crucial for the success of the prototype. The need for developing the prototype in an
open source environment was also recognised. The development of the prototype in
open source technology would assist in easy integration of the algorithm within the breast
screening services. This strategy would not force the screening services on any additional
financial overheads for adopting this algorithm in the form of license fees or new software
tools.
Java based Attendance prediction by Artificial Intelligence for Breast
Screening (JAABS) algorithm
“JAABS” is the new algorithm designed and developed in JAVA ® based technology. All
the screening services has to have is a Java Runtime Environment (JRE) (which is free to
download) for running the algorithm (Griswold and Lawrence 2004).
As the design process was based on more of an evolutionary type, a modular design
strategy was selected. This would assist in parallel development of the implementation and
also enable testing as modules rather than one single monolithic type of programme. The
modular design also ensured that any additions or changes can be implemented without
affecting the other modules. This was possible by the Java’s Object Oriented technology
(Griswold and Lawrence 2004).
JAABS design
The “Artificial Intelligence Module” encompasses the data normaliser; the neural networks
and the results collator (Figure 30). As its predecessor (AI-ATT), the Java based algorithm
implements two different neural networks, feed forward Back-propagation Prunning Neural
Network (BPNN) and Radial Basis Function Neural network (RBFN). The neural networks
proposed needs the input data vector classified as binary values. This conversion of data
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
53
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
into binary is termed as normalisation. Each possible outcome in the input data is
represented as a binary value, a “1” indicates the presence of the value and alternatively “0”
represents the absence.
Figure 30 - AI module (JAABS algorithm implementation)
The basic neural network and its back propagation-based learning technique are common
for the two networks (BPNN & RBFN). The input data in the RBFN is first passed through a
radial basis function algorithm, to identify the clusters and assign a radius for cluster
classification. These cluster centres are calculated and the real time data is checked against
these established clusters centres. The distance of the data point vector or record within the
input data space is arbitrarily assigned to the nearest cluster group. This data is then fed to
the neural network for making the prediction of attendance. Each episode has a different set
of predictor attributes; hence each episode is fed through separate neural networks which
were trained with their respective training dataset (Negnevitsky 2004, Haykin 1999,
Wasserman 1993, MacClelland and Rumelhart 1986, Bigus and Bigus 2001, Anderson 1995,
Masters 1995, Hudson and Cohen 2000, Moor 2003, Joos 2005).
Each dataset classified earlier into those with Townsend reference number and without (this
data set has the “Postannum” as the predictor variable) is used as the input for the
normaliser. This normalised data set is fed into the stream to generate the predictions. All
the results, one each from the Townsend reference and “Postannum” number stream of
output from the BPNN and RBFN network are collated by the “Results collator”. The
results module collects the collated prediction for each episode and submits to a “Poller”
based classifier. The “Poller” finds the best prediction for the given episode, for each stream
of dataset with Townsend reference number or without. The “Poller” generates the final
prediction output based on the confidence value of the prediction. This is fed in to the
prediction results collator for all the input (women) based on each episode. This is
associated with the women SX number and NHS number. The consolidated result is used to
generate the non-attendance list and written as a flat file for processing by the “Messaging
Module” for message generation.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
54
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
JAABS classes
The JAABS algorithm is designed as shown in Figure 31. The interface for the whole
algorithm is through “JAABSAlgorithmInterface” class. This class provides the sequential
logic for implementing the feed forward back propagation pruning neural network and the
radial basis function neural network. This module also utilises the IO module (sub module)
of the Data Pre-processing Module for input and output functions on dataset. The
“JAABSAlgorithmInterface” class instantiates two normalising classes, one each for BPNN
and RBFN respectively. The normalisation objects accept input file name and output file
name as parameters. This generates the collated predictor attributes with binary values for
feeding into the neural networks.
Figure 31 - JAABS algorithm (class diagram)
The BPNN neural network is implemented through the “BPNNAlgorithm” class. This class
works on a multilayer feed forward model. This class is instantiated by passing the
following parameters; number of input nodes, hidden nodes, output nodes and values for
learning rate, momentum and tolerance for training the network respectively. The
multilayer feed forward network refers to the predictor variables passing through the input
nodes, hidden nodes and to the output node as a stream. During the training mode, a
convergence strategy is adopted. In this strategy, a small proportion of the error value is
back propagated until the network error is within specified limits or if it completes the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
55
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
number of specified epochs (training cycles). After the training is completed, the network is
exposed to the validation data set (refer to later sections for details of training, validation,
and test dataset generation) and the predictions are saved as a flat file. The same way the
final testing set (data not exposed during the training and validating sequence) is fed
through the network to get the prediction of previously unseen dataset by the network.
The JAABS interface sequentially instantiates the RBFN stream normalisation and feeds the
binary data set to the “RBFNAlgorithm” class as “DataPoint” objects. This class instantiates
the input data as data vectors which is associated with a cluster through “Cluster” object. A
“Radius” object is created to refer to the distance of the input data vector to the cluster
centre. A “Centroid” object refers to the centres of each cluster. The complete activities and
their interrelationships of the JAABS algorithm stream are shown in Figure 32. The
additional episodes and the extensibility to any number of maximum episodes are clearly
indicated. The entire diagram provides an understanding of the required degree of
automation that has been incorporated within the JAABS algorithm.
Testing JAABS algorithm with simulated dataset
The initial testing of the JAABS algorithm was not tested with real-time data. A simulated
data set was created which contained values that can be mapped to the input domain. This
testing is aimed only at verifying the functioning of the JAABS algorithm and is not
intended to check the efficiency of the algorithm. This testing generated attributes based on
the Receiver Operating Characteristics (ROC) analysis (Lavra et al. 2003).
Figure 32- JAABS algorithm
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
56
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The ROC analysis yielded the figures quoted in Table 5 for True Positive Rate (TPR), False
Negative Rate (FNR), True Negative Rate (TNR), False Positive Rate (FPR), Positive
Predictive Value (PPV) and Negative Predictive Value (NPV) (Figure 33). Based on the
above said metrics further classifier characteristics can be evaluated, like Macro Average
(MA), Break Even (BE) and F-Measure (FM) (Lavra et al. 2003).
ROC
Metrics
TPR
FNR
TNR
FPR
PPV
NPV
MV
BE
FM
BPNN Network
Training Validatin Testing
set
g set
set
100.00%
98.25% 93.70%
6.30%
0.00%
1.75%
4.17%
100.00%
83.33%
0.00%
16.67% 95.83%
100.00%
80.00% 93.43%
4.35%
100.00%
98.59%
100.00%
90.79% 48.93%
100.00%
89.12% 93.56%
0.29%
0.00%
1.94%
RBFN Network
Training Validatin Testing
set
g set
set
98.18%
93.75% 94.07%
1.82%
6.25%
5.93%
81.18%
56.00% 41.18%
18.82%
44.00% 58.82%
77.14%
21.43% 96.21%
98.57%
98.59% 30.43%
89.68%
74.88% 67.63%
87.66%
57.59% 95.14%
2.04%
10.70%
1.90%
Table 5 - JAABS Algorithm's ROC metrics
100%
80%
60%
40%
20%
0%
TPR
FNR
TNR
FPR
PPV
NPV
MV
BE
FM
BPNN Network Training set
BPNN Network Validating set
BPNN Network Testing set
RBFN Network Training set
RBFN Network Validating set
RBFN Network Testing set
Figure 33 - The ROC characteristics for the JAABS algorithm
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
57
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The Data Pre-processing module processed the flat file and fed the data to the JAABS
algorithm (Figure 32). The data within the algorithm was first normalised and fed to the
BPNN and RBFN network respectively and prediction from these networks were written in
to the output file. The data from the prediction has been analysed in MS EXCEL® to
generate the ROC characteristics. The TPR, TNR, PPV, NPV, MV and BE metrics for the
BPNN was the maximum for the training set since the network has already been exposed to
this dataset. Whereas, the metrics for RBFN training set was not 100% as in the BPNN
training set metrics. For the current project, the most important metric in the ROC analyses
is the TNR. This metric corresponds to the prediction rate of the non-attendance. The
BPNN network TNR rate was only 4% (1) for the test set (Figure 34). The RBFN reported a
better TNR for the same test set at forty one percent (41%).
RBFN Network
BPNN Network
TNR
(7)
FPR
(23)
TNR
(1)
FPR
(10)
FNR
(16)
FNR
(22)
TPR
(327)
TPR
FNR
TNR
TPR
(254)
FPR
TPR
FNR
TNR
FPR
Figure 34 - ROC values on the predicted women (in numbers) for BPNN and RBFN network
These figures are nowhere near a perfect prediction metric; the reason can be mainly
attributed to the fact that the input dataset was a simulated one and the majority of the cases
being positive (women that attend), therefore the algorithm learns those cases that have
more probability to be true and if not tuned properly, almost does not identify the other
cases (non-attendance) at all. The dataset may not represent the a real-time domain values
so that prediction can be achieved with reasonable accuracy. Where as the TPR values
(Figure 34) are very high, these values are related to the positive prediction of the women
who are probable attendees. The RBFN network has a relatively higher TPR rate. BPNN’s
TPR is not far behind. This demonstrates that the planned classifier to identify the relatively
better prediction outcome among the two networks based on the confidence value can be
achieved than any of the algorithms used independently. Also this is in concurrence with
the strategy adopted in the predecessor’s (AI-ATT) work for employing such a classifier.
Further analysis with a real time data set is described in the later sections of the current
chapter.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
58
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Comparison of AI-ATT with JAABS implementation
One of the current project’s deliverables is the prototype implementation through
the JAABS algorithm. It is imperative that this algorithm be compared with the
previous AI-ATT algorithm to identify performance changes (perhaps due to
improvements of the JAABS algorithm). This strategy was adopted based on the
information collected during the qualitative analysis (open ended questions to the
experts and semi structured interviews). The AI-ATT algorithm was designed and
validated on the dataset downloaded from the Warwickshire, Solihull and Coventry
Breast Screening Unit (Arochena 2003).. The dataset was more than five years old and
lot of transformations had occurred within the same breast screening unit. The
upper age limit for routine screening has been increased to seventy (Patnick 2006).
This has resulted in more episodes than the AI-ATT algorithm has considered. The
maximum number of episodes used for attendance prediction by AI-ATT was four.
The current research has to take into account the possibility of an increased number
of episodes, the normal maximum number of episodes for JAABS algorithm has
been expected to be seven. The JAABS algorithm can easily be extended to cater to
any number of episodes for a given female population. During the qualitative research
stage, the breast screening unit’s manager had made a suggestion to identify more
predictors to improve the efficacy of the new algorithm. One major contribution to the
prediction algorithm made by JAABS in this project work is to identify more attributes.
There were several restrictions recognized in this initiative:
•
•
•
•
•
•
The new attribute should be deduced from the screening unit’s database
Data security, data integrity and privacy issues should not be breached
If any processing is mandatory for the new variable, then it should be viable
with existing computing technologies
There should not be put undue resource constraints on the screening unit
The coding should tie in to the JAABS algorithm
The independency (non-correlation) hypothesis should be satisfied with the
already selected predictors for accurate use of NN
These restrictions made the selection of new predictor variable very complex. Further
investigation on the dataset, revealed a possible new attribute in the form of distance
between the women’s residence and screening unit’s location for taking the mammogram.
The qualitative research established that the postcode of the women’s residence and the
postcode of the screening unit can be used to calculate the required distance variable. This
distance factor was identified as a predictor after due exploratory tests and later included as
one of the predictor attributes in the JAABS algorithm. The validation of this new attribute
and the relative performance increase by the new attribute to the screening prediction is
enumerated in the later sections. During the qualitative research stage, the breast screening
unit’s manager had suggested the identification of more predictors to improve the efficacy
of the new algorithm. Further investigation on the dataset, revealed a possible new attribute
in the form of distance between the women’s residence and screening unit’s location for
taking the mammogram.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
59
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The qualitative research established that the postcode of the women’s residence and the
postcode of the screening unit can be used to calculate the required distance variable. In
addition to this it was also found that the screening unit was at different locations and the
mobile units gave the portability factor to the screening equipment. The clinic code was
identified as a possible constituent for processing the new distance-based variable. A novel
approach of using area code (first part of the postcode) related to the residence and the
screening unit (imaging location) was proposed. The “ClinicCode” represented the location
of the mobile or stationary units. The screening unit’s postcode (area code) and the
women’s postcode (area code) were separated and provided as parameters to the
“DistanceCalculator” class to calculate the distance.
The “DistanceCalculator” class matched the area codes to their respective latitudes and
longitudes from a lookup file and then the distance was calculated in miles/kilometres. This
distance was later categorised as a new variable and named as “Scr dist”. This new
distance-based variable was not part of the AI-ATT algorithm and was recommended as one
of the future predictor attributes.
JAABS evaluation
The JAABS algorithm was looked upon as an opportunity to maintain the continuity of the
earlier exploratory research and for validating the same. Hence JAABS was tested with the
data collected from the Warwickshire, Solihull and Coventry Breast Screening Unit, which was
also the data source for validating the AI-ATT algorithm. The screening unit manager
insisted to pre-process the data within the screening unit’s Local Area Network (LAN) to
ensure data protection and data security. The pre-processor module was executed in Java ™
2 Runtime Environment (JRE), standard edition (build 1.4.2_04-b05) version 1.4.2_04. The
pre-processor module was designed based on object oriented concepts. This assisted in
easily detaching the module from the JAABS algorithm and ported to the screening unit’s
LAN. The testing of JAABS algorithm on the simulated dataset ensured that without any
modifications to the code, the pre-processing module did not encounter difficulties when
presented with the real-time dataset.
Dataset
Prior to data extraction, the Crystal Reports (CR) template was first tested with a test run.
Refer to Figure 35 for the condition used to select client Sx number in the “Select Expert”
dialog window, an arbitrary dataset comprising women’s record between 100001 to 250000
(Sx number) was used to extract a test set. After manually checking, it was found that the
data pertaining to the period prior to 1995 was not complete (especially clinic code) and
hence could not be used in the pre-processor module, so the strategy was altered and
instead of using the Sx number as the selection criteria, women episode’s date of first offered
appointment was selected (Figure 36).
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
60
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 35 - Test run for CR template verification
The template included all previously screened women’s records between the dates 01
January 1995 to 01 January 2008.
Figure 36 - Selection criteria used for the entire dataset
The CR template for this extraction used the same data structure as depicted in Figure 37.
The total number of records generated by this data extracts using the final version of the
CR template was 540,539. This extract in text form was used as the input file for the data
pre-processing module. The programme took 227 minutes to complete the process of
generating new attributes, filtering and sorting for the whole record extract previously
generated by the CR template.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
61
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Figure 37 - Final version of CR template
This module identified episodes with missing data and removed them from the study. In
total 2% (9,799) were removed as records with missing data (Table 6). It further deleted
almost 3% (15,778) of the total records due to duplicate entries. The valid records constituted
86% (159,405) of the extracted dataset on an average each record had 3.2 episodes.
Description
Total valid women’s record
Number of records deleted due to
multiple entries
Records with missing values
CR template output records
Number of records
159,405
15,778
9,799
540,539
Table 6 - Records breakdown details after data pre-processor module
The pie chart (Figure 38) shows the actual layout of the input dataset after classification by
the pre-processor module. These valid records were associated with demographic details
and then categorised based on the number of episodes each women’s record contained.
Based on this classification, eight episode groups were formed. Episode-wise details are
listed in Table 7. The eighth episode group had only 16 records and was too small for
training and testing, hence has been omitted from further analyses.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
62
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Number of records
deleted due to
multiple entries,
15778, 3%
Records w ith
missing values,
9799, 2%
Total records
before collation,
514962, 95%
Figure 38 - Pie-chart showing records classification after pre-processing
Table 7 depicts the spread of data for each episode. The highest number of records was
reached for the fourth episode. The first to fifth episodes had an average of 31,000 records.
For the remaining episodes (6th, 7th and 8th) the average is only 800 records, this might have
strong impact on the actual prediction capacity of the JAABS algorithm for these episodes.
Episode number
Episode 1
Episode 2
Episode 3
Episode 4
Episode 5
Episode 6
Episode 7
Episode 8
Total records
23270
33765
29497
43584
26669
2366
238
16
Train set
4646
6838
5868
8792
5340
473
36
NA
Valid set
4708
6734
5891
8839
5338
485
39
NA
Test set
13916
20193
17738
25953
15991
1408
163
NA
Table 7 - Training, validation and test sets classified by the pre-processor module
The screening programme has recently initiated the increase in the upper age limit from 64
to 70 years. These screening batch records are in their infancy, hence the lower number of
historical records for these episodes. On an average the call/recall cycle takes three years to
complete a repeat of the whole population in the screening units care delivery area. This
fourteen year period can be mapped to a maximum of five episodes. This was the case for
the previous work carried out with the AI-ATT algorithm; it was designed for only four
episodes, as a dataset was only available for those episodes. The latest increase in the age
limit has certainly had a huge impact on the number of processable screening episodes. This
can be verified in the current dataset, the fifth episode has 26,669 records.
This trend for episode-wise population increase is expected to continue as the number of
women attending the later episodes from fifth to eighth also increases. This clearly signifies
the importance of modifying the JAABS algorithm to accommodate more episodes, catering
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
63
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
to the future screening unit’s requirements. The next step is to tri-furcate the dataset into
training, validation, and testing sets. For each episode, a 20:20:60 proportion (for training,
validation and testing respectively) was utilized to generate a complete suite of dataset
(Figure 39). The training and validating sets were generated randomly by SPSS software.
The remaining data (about 60%) constituted the test set. Such a classification was
compulsory, ensuring a balanced approach to predicting the outcome attribute. The
balanced set will contain equal amount of attendees and non-attendees. The remaining data
was classified as a test set that had an unbalanced dataset i.e. the non-attendees records were
only a fraction of the attendees. The future datasets which the algorithm has to predict in
real-time would have an un-balanced dataset similar to the test set. The composition of the
datasets thus created is depicted in Fugire 39. The tri-furcated dataset was saved as separate
text files with a pre-determined name for further processing by the prediction module.
45000
40000
Number of records
35000
30000
25000
20000
15000
10000
5000
0
Episode 1 Episode 2 Episode 3 Episode 4 Episode 5 Episode 6 Episode 7
Train set
Valid set
Test set
Figure 39 - Tri-furcation of dataset into training, validation and testing set for seven episodes
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
64
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Attribute selection
The next stage of the validation sequence was complete analyses of the dataset and its
association with the prediction attribute i.e. screening attendance. The dataset attributes of
the AI-ATT was analysed and it was found that each episode had different sets of predictor
variables (Table 8).. The predictor variables being categorical were analysed through
parameters such as Lambda, Uncertainty, Phi (), Crammer’s V and Contingency
(confidence level at 95%). These tests were conducted on the probable predictor’s
(independent variable) association to the screening attendance (dependent variable). Table 9
summarises the degrees of association of the independent variable to the screening
attendance (highest value among the tests). The AI-ATT predictor attributes were tested on
the new dataset. Upon analysing, the results indicated that some form of association,
however weak can contribute to better classifying of the outcome.
Episode
numbers
EPISODE 1
EPISODE 2
EPISODE 3
EPISODE 4
ANNP
(Post
annum)
Age band-1
Type bin-1
Type bin-2
Post annum
Att bin-1
Age band-2
Cancer -2
RBFN
(Post
annum)
Age band-1
Type bin-1
Att bin-1
Age band-2
Cancer-1
Post annum
Type bin-2
Num test-1
FP-1
ANNP
(Townsend
reference)
Age band-1
Type bin-1
Cancer-1
Att bin-1
Type bin-2
Int slip-1
Age band-2
Post annum
RBFN
(Townsend
reference)
Age band-1
Type bin-1
Att bin-1
Cancer-1
FP-1
Num test-1
Age band-2
Int slip-1
Att bin-2
Cancer-2
Hist FP-2
Age band-3
FP-2
Att bin-3
Age band-4
FP-3
Hist FP-3
FP-2
Hist can-2
Att bin-2
Cancer-2
Att bin-2
Num test-2
Att bin-2
Att bin-3
Cancer-3
Age band-4
FP-3
Type bin-4
Att bin-3
Cancer-3
Age band-4
Int slip-3
Att bin-3
Cancer-3
Hist FP-3
Age band-4
Table 8 - Predictor attributes for AI-ATT algorithm
It is to be noticed that, episodes 6 to 8 have lesser records for training, validating and testing
and they were not given more weightage when deciding the predictor variables for JAABS.
The rest of the episodes 2 to 5 indicated that all the variables have varying degree of
association with the screening attendance. For episode 1, the same set of attributes as those
of AI-ATT were used for current validation on JAABS. The new predictor variable in the
form of screening distance was also tested for its association with the screening attendance
attribute. For almost all the episodes varying degrees of associations were acknowledged.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
65
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Hence further detailed analysis was crucial to commit to the decision, whether to include
these attributes as predictors or not.
Independent
variables
Epi1
Epi2
Epi3
Epi4
Epi5
Epi6
Epi7
Epi8
PA
TS
AttBin
NumTest
Cancer
FP
HFP
HC
AttTypeBin
AgeBand
Slip
ScrDist
Variables used in AI-ATT
Association more than 0.2
Association more than 0.1 and less than 0.2
Association more than 0.0 and less than 0.1
Association less than 0.0
Table 9 - Variables association level with screening attendance for the new dataset
These issues were fundamental and are to be resolved before the entire dataset can be
validated. The second episode was selected for trying out different strategies that may be
adopted for the JAABS algorithm. The first part of such a strategy will decide what predictor
attributes will be employed for JAABS algorithm. The other strategy will verify the validity
of using screening distance as one of the predictor attributes.
JAABS vs AI-ATT attributes
The dataset for episode two was employed for training, validating and testing the JAABS
algorithm on two different set of predictor attributes. One set consisted of variables
employed in AI-ATT (Table 8). The second set contained the entire predictor variables (post
annum-PA, Townsend reference-TS, attendance of previous episode- Att Bin, number of
tests conducted in previous episode-Num test, cancer detection in previous episode- Cancer,
false positive in previous episode-FP, history of false positive-HFP, history of cancer
detected in earlier episodes-HC, invitation type for current episode-Att Type Bin, age band
for the current episode-Age band, number of days in difference with actual attendance to
screening appointment date-Slip and screening distance travelled by the women from her
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
66
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
residence to screening unit- Scr Dist).. Since the degrees of association for second episode
was similar to the episodes three to five and the total number of records for the second
episode were approximately equal to the rest of the episodes (third to fifth episode), hence
the second episode was identified to test these strategies.
Test
Valid
Train
Type
JAABS with all predictor variables
JAABS with AI-ATT variables
JAABS with all predictor variables
JAABS with AI-ATT variables
JAABS with all predictor variables
ACC
69.3
69.6
69
71.9
100
NPV
47.6
48
30.2
32.7
100
PPV
77.8
78.1
81.3
83
100
SPC
45.6
46
33.9
35.2
100
SEN
79.1
79.4
78.6
81.4
100
JAABS with AI-ATT variables
100
100
100
100
100
Table 10 - Average ROC characteristics for JAABS with AI-ATT variables
The training set had 100% ROC values for both the attribute sets (Table 10). In the validation
set the ROC characteristics for AI-ATT variable alone were little better (nearly 3% on
accuracy and 3% on negative predictive value were reported).. Where as the test set for both
the set of variables had almost the same ROC characteristics. Hence the strategy of using all
the predictor variables did not affect the prediction characteristics; hence this strategy was
employed for the rest of the episodes, except for episode one in JAABS algorithm. This
strategy had its own advantages in the form of easy coding, less maintenance and effortless
extensibility of the algorithm in catering to future extension of screening population age
limits. Further, the normalisation and the classifier can be coded as common modules for all
the episodes.
Screening distance as a new predictor attribute
Set
Test
Valid
Train
Type
With ScrDist
Variable
Without
With ScrDist
Variable
Without
With ScrDist
Variable
Without
ACC
JAABS
NPV
PPV
SPC
SEN
69.32105
47.61318
77.81069
45.62795
79.15761
67.37483
44.83248
76.11325
42.1149
78.06597
68.98901
30.20566
81.29717
33.88609
78.58833
64.49361
27.19715
76.93069
28.21935
76.01252
100
100
100
100
100
100
100
100
100
100
Table 11 - Screening distance test ROC values
The screening distance variable was included as a predictor in one set of tests and a second
set without. The training data had been exposed to the algorithm hence had 100% ROC
values. The validation set with screening distance attribute indicated a 5% increase to ACC,
3% on NPV, 4% on PPV, 5% on SPC and 2.5% on SEN. The difference between the ROC
characteristics for the test set was from 2 to 3% (Table 11). This indicated that better
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
67
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
prediction can be achieved with screening distance variable, but further detailed studies can
assist in fine tuning the JAABS algorithm.
JAABS ROC tests
The earlier sections described the strategy adopted for devising the constituents of the
predictor attributes. This section discusses the ROC characteristics of the tests conducted on
the JAABS algorithm with the identified set of predictors. The dataset created after trifurcation into training, validation and test sets were applied on the JAABS algorithm. The
training and testing was conducted for ANNP and RBFN networks as separate activities.
Finally the fully automated JAABS algorithm with the voting-based classifier was tested and
the results are listed in this section.
RBFN test
The RBFN stream was presented first with the training set and then validated with the
balanced validated set and finally tested with an unbalanced test set, simulating the realtime prediction data. The only difference between this test and the real time prediction set
would be a lower number of women in a given batch. For episode one, the lower number of
attributes (due to the lack of screening history) and the approximation capabilities the RBFN
network had almost similar characteristics with the training, validation and test sets. Nearly
47% of correct prediction (CP) and 47% of False Negative (FN) were reported (Figure 40).
Episode two had improved values, CP had 67% and FN decreased to 12% and False Positive
(FP) at 21%. The CP, FP and FN were found to be consistent around these same values for
episodes two to five. Incidentally the number of training records was higher in these
episodes. This may answer the increase in accuracy of CP for these datasets. Episode six has
poor predicting capabilities, the CP was at 53%, FP at 5% and FN at 42% for the test set.
These figures were consistent with the training and validating set, these could be attributed
to insufficient training data. Episode seven had similar values (CP 67%, FP 7% and FN 26%
for validation and CP 83%, FP 3% and FN 14%) for training, but these results were
inconclusive due to the lower training and testing data.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
68
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Test set
Validation set
FP
6%
Training set
FP
6%
CP
47%
CP
46%
FN
47%
FP
21%
CP
67%
FN
12%
Episode 2
Test set
CP
46%
Episode 1
Episode 1
Test set
FP
6%
FP
19%
Validation set
FN
48%
FP
21%
FN
12%
CP
67%
Episode 2
Validation set
FP
18%
Episode 1
Training set
CP
69%
Episode 2
Training set
FN
48%
FP
20%
FN
11%
FP
19%
FN
10%
CP
70%
Episode 3
Test set
FN
11%
CP
71%
Episode 3
Validation set
FP
16%
FN
11%
FP
16%
CP
71%
Episode 3
Training set
FP
16%
FN
11%
CP
73%
Episode 4
CP
73%
Episode 4
FN
11%
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
CP
73%
FN
11%
Episode 4
69
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Test set
CP
69%
Episode 5
Test set
CP
51%
FP
20%
FN
11%
FP
11%
FN
38%
Episode 7
Test set
CP
53%
Episode 6
FP
5%
FN
42%
Validation set
FP
20%
CP
69%
Episode 5
Validation set
FN
11%
FP
7%
CP
67%
Episode 7
Validation set
CP
53%
Episode 6
CP
69%
Episode 5
Training set
FN
26%
FP
7%
FN
11%
FP
3%
FN
14%
CP
83%
Episode 7
Training set
FN
40%
FP
20%
Training set
CP
54%
Episode 6
FP
6%
FN
40%
Figure 40 - RBFN stream prediction details for all episodes
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
70
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
ANNP test
The ANNP stream, part of the prediction algorithm was similarly tested with the trifurcated dataset (Figure 41).
Test set
CP
67%
Episode 1
Test set
FP
17%
FN
16%
FP
16%FN
15%
Test set
CP
65%
FN
19%
Validation set
FP
14%
FN
16%
CP
70%
Episode 2
Episode 2
CP
71%
Episode 3
FP
16%
Episode 1
CP
69%
Test set
Validation set
FP
15%
FN
14%
Validation set
FP
12%
FN
14%
CP
74%
Episode 3
FP
11% FN
Validation set
10%
CP
79%
CP
80%
Episode 4
Episode 4
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
FP
9%
FN
11%
71
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Test set
Validation set
FP
12%
FN
13%
FN
11%
CP
76%
CP
77%
Episode 5
Test set
CP
58%
Episode 6
Test set
FP
11%
Episode 5
FP
22%
FN
20%
FP
17%
Validation set
CP
59%
Episode 6
Validation set
FP
19%
FN
22%
FP
5%
FN
23%
CP
59%
Episode 7
FN
24%
CP
72%
Episode 7
Figure 41 - Prediction characteristics for ANNP
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
72
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The training set was always found to be 100% correct on CP, FN and FP, hence it is not
represented in the figure. Similar to the RBFN test values, ANNP also had poor first episode
prediction characteristics. Yet the FN was less than one third (16%) of the RBFN stream. The
FP was three times (16%) of the RBFN value for the testing set. The CP was the highest at
67%. The validating set had more or less similar values. Episode two and three also had
typical values, but they were improving steadily. Episode four with the maximum training
data had the best values of CP, FP and FN, 79%, 10% and 11% respectively. Episode five also
had similar but were less than episode four values. Like RBFN stream, ANNP also had
similar prediction values. The CP was around 60%, this may be attributed to the lesser
training dataset. The marked difference was seen in the first episode, in comparison, ANNP
was better predicting in spite of the lower number of predictor variables (ANNP-67% vs
RBFN- 47%).
JAABS test
The JAABS algorithm was tested with the complete set of episodes after appropriate training
and validating. The ROC characteristics are summarised in Table 12. The algorithm’s final
prediction of the screening attendance was based on a polling strategy that relies on the
prediction confidence. The accuracy of the algorithm was around 68% for the first three
episodes. Episode four had the maximum accuracy at 79% closely followed by the fifth
episode. The accuracy of the sixth and the seventh were the lowest between 51% and 57%.
The NPV was the maximum at 51% for the fifth episode. The rest of the episodes had NPV
values between 41% and 47%.
JAABS
Test
ACC
NPV
PPV
SPC
SEN
Episode 1
67.29
42.07
76.71
40.22
78.05
Episode 2
69.38
47.65
77.87
45.66
79.22
Episode 3
69.95
39.45
76.46
26.29
85.59
Episode 4
79.17
39.25
87.06
37.37
87.93
Episode 5
76.23
51.61
83.84
49.64
84.89
Episode 6
57.79
46.51
64.77
44.92
66.21
Episode 7
51.39
30.02
76.53
60.05
48.18
Average for 4 episodes
71.45
42.11
79.53
37.39
82.7
Average for all episodes
67.31
42.37
77.61
43.45
75.72
Table 12 - ROC characteristics for JAABS algorithm
Episode seven had the lowest NPV (30%). These lower NPVs were expected as the
proportion of non-attenders was lesser in the test set. The PPV for the fourth and fifth
episodes were higher between 83%-87%. The remaining episodes had values in the range of
seventies, except for sixth episode; it was 64%. The specificity metric is the most important
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
73
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
characteristic of the ROC values. Specificity indicates the true negatives in the population;
this is related to the screening non-attenders for the given test batch. Specificity was the
highest for the seventh episode at 60%, but this may not be a true indicator as this episode
had only 238 records in total. The next highest value was in the fifth episode at 49%. Episode
one, two and sixth had values between 40% and 45%. Episode 3 and four had lower values
at 26 and 37% respectively. The sensitivity was around 80% for the first four episodes,
peaking at 85%for episode 3. The higher the training set of records, higher the sensitivity
values. The average for the first four episodes can be used for comparing the JAABS and AIATT algorithms. The ROC values for AI-ATT algorithm is shown in Table 13.
AI-ATT
Test
ACC
NPV
PPV
SPC
SEN
Episode 1
67.01
20.45
87.48
41.81
71.43
Episode 2
87.76
56.1
92.85
58.91
93.14
Episode 3
86.49
50.54
92.91
55.99
91.32
Episode 4
81.65
41.26
92.51
64.59
85.42
AI-ATT Avg
80.73
42.09
91.44
55.33
85.33
Table 13 - ROC characteristics of AI-ATT algorithm
Figure 42 shows the comparison of the average ROC values for the first four episodes for
the AI-ATT and JAABS algorithm. Even though all the ROC characteristics were better for
AI-ATT, the negative predictive values were almost the same at 42% for both the algorithms.
Being an exploratory research and the first iteration of a prototype, JAABS has enough
potential for acceptance into the routine activities of screening unit. Undoubtedly a number
of upgrades have been identified. One such area for improvement is the neural net’s ROC
characteristics for the new set of data. An important facet that could be part of future work is
to improve the specificity of the JAABS algorithm. Figure 42 shows the difference in the
average ROC values between the algorithms episode wise. The main handicap identified for
the AI-ATT algorithm was the prediction capacity of the first episode. Incidentally the
JAABS algorithms had better prediction values for this episode and this not only suggest
that the characteristics of neural network-based prediction is working better with lower
number of predictor variables but also suggests that as the input dataset evolves and when
enough training sets are available would definitely assist in making decent screening
attendance predictions.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
74
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
100
91.44
90
80
85.33
82.7
79.53
80.73
71.45
70
55.33
60
50
42.09 42.11
37.39
40
30
20
10
0
ACC
AIATT Avg
NPV
PPV
SPC
SEN
JAABS average for first 4 Episodes
Figure 42 - JAABS vs AI-ATT average ROC values for first four episodes
Summary
Even though the tests conducted to recognize the association between the independent and
dependent variables showed varied degrees of strengths. These tests for association are only
an indication of a linear relationship, whereas the real problem space is multi-spatial. Hence
a new strategy of exploiting all the available predictor variables was proposed and validated
in the JAABS algorithm. The accuracy and sensitivity parameters exhibit the predicting
capacity of the models (Table 12 and Table 13). The values are not significantly different
between the two algorithms. Based on the quantitative analysis performed on the output
parameters and the earlier algorithm’s (AI-ATT) ROC values, triangulation techniques were
used to compare JAABS, AI-ATT and Logistic Regression (LR) models. It was found that for
episodes one to four, the LR model had only 9% NPV, where as the other two algorithms
had 42% NPV for the same set of episodes. Especially for the first episode, JAABS had a
better NPV over AI-ATT (42% vs 20%). From the model’s performance perspective, all these
prediction characteristics were positive.
The AI model (JAABS) proposed were consistent to the earlier model’s (AI-ATT)
performance and consistently out-performed traditional statistical techniques (LR). This
could be attributed to the complex nature of the input variables. The knowledge creation by
applying AI is not only consistent, repeatable and economical; but also ensures minimal
human intervention. This is perfectly ideal for automating the whole process. These were the
principal requirements for the prototype model that has been fulfilled through appropriate
validation in this chapter.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
75
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Knowledge Sharing: GP’s Role
In the last chapter, the GP’s role and the proposed knowledge sharing were described in
detail. This chapter identifies the importance of GP interventions and their impact on breast
screening attendance. Earlier chapters described how quantitative research would be
applied through a questionnaire-based study. In concurrence to this, the following
discussion first covers the factors which influence the womens’ breast screening attendance
and details that are related to approaches that can be used to address such factors. It
provides further analyses through a questionnaire-based survey conducted with Coventry,
Solihull and Warwickshire GPs. The survey covered a wide range of areas related to
interventions; including the favoured types of interventions, mode of message delivery,
updating methods and requirement of additional resources. Combined with the information
collected during the pilot study (from the GPs), this analysis would shed more light on the
finer points of intervention and assist to triangulate these results.
GP intervention
Breast screening non-attendance can be attributed to semi-permanent and temporary factors.
Semi-permanent factors such as ethnicity, age group, marital status, income, education and
chronic conditions can be major influencing factors on the screening population (Katz et al.
2000). These factors are difficult to address and usually consume more time and resources
to make a significant impact; whereas temporary factors such as employment, personal
apprehensions, beliefs, knowledge and access to screening facilities can be targeted
effectively to increase the screening attendance (Sin and Leger 1999, Bekker et al. 1999)..
Publicity campaigns, media-based advertisements, knowledge-based education drives,
posters, pamphlets etc. are the primary counter measures planned and executed by the NHS
to address the temporary factors affecting screening attendance. Research by Kelley has
found that by bringing the screening services to the women rather than making the women
come to the screening office can greatly influence the attendance rate (Day et al. 1989). The
NHS has many mobile screening units under its service and they are strategically placed at
public places such as shopping centres and other public frequented places for maximum
effectiveness.
Types of intervention
Evidence points out that intervention through printed and audio-visual educational
materials and education sessions has limited effectiveness, but when these strategies are
coupled with GP interventions, they can help the non-attending women to make informed
decisions on screening (Jepson et al. 2000). Primary care can play a vital role in addressing
some of the important temporary factors affecting screening attendance such as personal
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
76
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
apprehensions, beliefs, and knowledge on breast cancer screening (Fox et al. 1991, Bekker et
al. 1999).. Several studies have established that primary care physician’s (GP’s) advice and
recommendations are important factors that guide women to make educated decisions on
screening (Sin and Leger 1999, Kelley 1999, Bennett et al. 2000, Vainio and Bianchini 2002,
Day et al. 1995, Turner et al. 1994, Fox et al. 1991). Further research has identified that family
physicians can considerably help to increase attendance through appropriate interventions.
These can take the form of GP signed letters, one-to-one meetings, telephonic or
opportunistic interventions (Turner et al. 1994).. The current research work proposed to
create knowledge on screening non-attendance through prediction and share it with
primary care deliverers through effective knowledge sharing techniques (Baskaran et al.
2004). Such knowledge when appropriately shared can be leveraged to initiate reminders at
the right time in a clinical context. Reminders, if properly utilised, could be part of a
potential strategy to improve the underused preventive care services (screening) (Shea et al.
1996). The lack of research in this area prompted to further investigate the intricacies and
strategies to be employed while adopting the proposed KM-based approach.
Questionnaire-based survey
The lack of evidence in this area and the recent advances in Information Technology (IT)
prompted the collection of additional data to determine the best possible use of the gained
knowledge, in particular data about the preferred mode of knowledge delivery to GPs, their
opinion on the interventions and their perceived need for additional resources/assistance
(refer to blank questionnaire in Error! Reference source not found. for more details). This
research conducted a survey of the local GP surgeries which are geographically related to a
local breast screening service in West Midlands, UK. All the GP surgeries (N=215) covered
by the Warwickshire, Solihull and Coventry Breast Screening Service were eligible for the
sample survey. Lists of GP surgeries were obtained from the area’s Primary Care trusts.
The GP surgery population spread across the service can be classified by the following areas:
Warwickshire (north) 38, Warwickshire (south) 48, Solihull 36, Rugby 15 and Coventry 78.
In total, the questionnaires were dispatched to 215 GP surgeries. Only 30 (14%) responses
were received in the first round and this was followed by a reminder letter which yielded an
additional 29 (13.5%) responses (refer to Error! Reference source not found.). A total
response rate of about 59 (28%) was achieved; this was slightly below average for a postal
questionnaire based survey. Three questionnaires were returned back due to change of
surgery address.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
77
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Sample
Domain
Screening
women
Description
Queries answered
Target population of the • Population (approx.) of women registered with GP
screening programme
• List of clinical condition and its frequency
Prediction
Providing
prior • usefulness of prediction
information
on
non- • appropriate duration for intervention
attendance
through
prediction
Electronic Data Electronic exchange of • Usefulness of paper based screening report
Interchange
information to GPs
• Frequency of screening report and its update to women’s
record
• Preferred mode of future knowledge and information delivery
Impact
Additional
information
Demography
Impact
of
GPs
screening attendance
on • Probable impact on breast screening attendance
•
•
to •
GP •
Aspects
related
recognition
of
intervention
Information on surgery •
and GP
•
•
•
Preferred intervention type and its duration
Preferred staff for intervention
Recognition for intervention
Requirement of additional resources
Sex
Experience
Number of partners
Surgery population
Table 14-Issues covered by the questionnaire (Baskaran et al. 2006)
Survey
The survey was conducted in the summer of 2006 through a mailed four page questionnaire
which enquired about GPs belief on the screening prediction, mode of knowledge delivery,
intentions with the proposed knowledge sharing and any additional resources required to
intervene (Table 14). The survey was accompanied by an executive summary of the research
envisaged. The survey design and analysis was approved by Coventry University’s
Research Ethics Committee and the NHS Local Research Ethics Committee (LREC). The
questionnaire was piloted for its context, logic and readability with two groups of experts,
one containing five academics within the university and five GPs located in the Coventry
area. The GPs suggested that the covering letter, if addressed to the practice manager,
would result in a better response rate (earlier it was addressed to the senior physician of the
surgery). Some of the questions were found too complex and the GPs suggested splitting
the questions into smaller and simpler ones for a better response. During the pilot, some of
the technical (non-clinical) terms were suggested to be avoided for better understanding.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
78
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Category
Screening
women
Prediction
EDI
Impact
Additional
information
Description
Dichotomised variable
Frequency of women approaching GP • Often
surgery
• not at all
Usefulness of prior information on non- • useful
attendance
• not useful
Usefulness of paper based screening • useful
results
• not useful
Helpfulness of electronic mode of results • preferred
delivery
• not preferred
Impact of GPs intervention on screening • weak impact
attendance
• strong impact
Preferred staff to conduct intervention
• preferred
• not preferred
Recognition recommended
• strongly
• weakly
Requirement of additional resources
• less
• more
Table 15- Dichotomisation of the variables categorised by questionnaire subsections
Earlier research by Bankhead et al. (2001) had suggested that intervention through GPs was
useful but was found not to be a cost effective strategy. Moreover, this study was conducted
before 2000 and the inferences (cost-based) may be outdated. Other observational studies
and manual prompts in women’s records were found to assist GPs in encouraging nonattending women to attend a scheduled screening appointment (Sin and Leger 1999).. The
current survey was expected to validate four important factors pertaining to the
intervention. The first assumption is that the screened women aged 50 to 70 would be
visiting the GPs on a regular basis. The next hypothesis relates to the opinion of GPs on the
vital part they can play in the quest for improving screening attendance through
intervention. The third assumption is that GPs would prefer opportunistic intervention
when compared to other types of intervention. The final hypothesis relates to the electronic
version of knowledge transfer and its preference over the traditional paper-based screening
results (Baskaran et al. 2006a).
Questionnaire analysis
The preliminary analyses of the responses pointed to a bimodal distribution. Hence the
responses are dichotomised as follows (refer to Table 15). The data was analysed using
SPSS© (12.0) for Windows software (Hinton et al. 2004). The attributes employed in the
analyses include demographic variables such as gender, practice experience, number of
partners and surgery population. The other attributes relate to screening women,
prediction, Electronic Data Interchange (EDI), GPs impact on attendance and information on
additional resources required (refer to Error! Reference source not found. for the results
frequency distribution).. Multi-variate and bi-variate analyses were conducted on the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
79
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
attributes to estimate and experiment the influence of the covariates on the various sets of
dichotomised variables. An exhaustive literature review within the study domain was
conducted to classify the appropriate attributes for this study. In addition, the independent
variables based on their importance were incorporated in the logistic regression model.
Variable
Usefulness of
screening
Usefulness of
Helpfulness of
results
prediction
electronic results
Frequenc
Frequenc
p
Frequen
p Value
y
p Value
y
Value
cy
0.131
Gender
Male
Female
Practice
experience
<10 years
>11 years
100%
(19)
89%
(32)
0.41
83% (15)
79% (26)
95% (35)
74% (28)
0.436
>2
<2
Registered
populatio
n
>5000
<5000
91%
(30)
96%
(24)
100%
(17)
81% (13)
78% (29)
81% (34)
74% (32)
0.215
91% (31)
76% (25)
92% (23)
80% (20)
77% (20)
0.914
0.663
69%
(11)
74%
(32)
0.917
71% (20)
0.943
93%
(26)
93%
(28)
0.741
72%
(13)
76%
(29)
0.583
0.053
88%
(14)
0.056
GPs ability to
intervene
Frequenc
p
y
Value
0.424
68% (13)
0.449
Partners
0.008
88% (15)
0.182
100%
(17)
90%
(37)
GP intervention
has strong impact
Frequenc
y
P Value
0.082
82%
(27)
62%
(16)
0.693
0.171
82% (18)
93% (26)
79% (22)
81% (25)
81% (25)
74% (23)
0.035
86%
(24)
61%
(19)
Table 16 - Cross tab analyses on screening, prediction, EDI and intervention against demographic
details
The multivariate and bi-variate analyses considered significant values of <0.05. Odds Ratio
(OR) and Confidence Interval (CI) were also performed. 38 (67%) of the respondents were
female GPs. 41 (73%) of the surgeries had four or less partners. Nearly 25 (49%) of the GPs
who responded had more than twenty one or more years of practice and 18 (35%) had
experience between eleven and twenty years. Nine respondents did not provide their
practice experience.
Screening women
The majority of responding surgeries 46 (81%) had a registered population of less than
10,000 and the remaining surgeries had populations of more than 10,000. The national
average for registered populations in UK surgeries is 8,400 (The Information Centre 2007),
whereas the average registered population for each practice within Coventry Primary Care
Trust PCT was 5000, Warwickshire PCT 4900 and for Solihull PCT was 6500 (The
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
80
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Information Centre 2007).Most of the respondent surgeries 40 (76%) had 400 or more women
(aged between 50 and 70) registered with them.
The majority of the respondents reported that women aged between 50 and 70 approached
the surgery more often for Muscular 46 (92%), Diabetes 44 (88%), Arthritis 44 (88%) and
Hypertension related clinical conditions 47 (94%). There was 10% missing data among the
respondents on this question. Nearly half, 26 (49%) of the responding GPs visited women
(aged between 50 and 70) at home for various health related services, 24 (37%) of nurses
made less than 5 visits per month to the same aged women. Special nurses made 15 (30%)
visits to less than 5 women per month for the same age group. The association between the
surgery demography and screening results usefulness has no statistical significance due to
higher p values, Table 16). .
Prediction
41 (82%) of the GPs agreed that it would be very useful to provide prior information
through prediction about non-attending women, so that they can intervene and reduce
breast screening non-attendance. The majority of the GPs, 50 (93%) are of the opinion that
about thirty days (in advance) would be an ideal period to initiate intervention on the
predicted non-attending women. The usefulness of the prediction was not significantly
associated with the surgery demographics, but those surgeries with less than two partners
were associated (to a certain extent) with the GPs who found prediction useful (Table 16)..
Electronic Data Interchange (EDI)
Most of the GPs 55 (93%) believed that it is very useful to receive screening results (in paper
format instead of electronic) to provide better health care to women. 52 (92%) GPs agree
that currently they are receiving screening results in paper format from the screening
service. Bivariate analyses of EDI and the gender of GPs who responded to the
questionnaire revealed that female GPs are more likely to accept electronic results (95% vs
68%, p=0.008). Almost the same percentage agreed that these results, after being received,
are always updated into the women’s records. Out of these GPs, 33 (83%) mentioned that
the results are manually read from the report and added into the women’s records. The
remaining surgeries answered that they are scanning the screening result document and
storing it as a computerised file along with the woman’s record in their local database. Even
though bivariate analyses did not show any significance that EDI based results are more
favoured by all GPs, further analyses indicated that GPs with less than ten years of
experience are more likely to accept electronic based screening results (100% vs 81%,
p=0.053). 51 (86%) of the GPs who responded to the questionnaire were of the opinion that it
would be very helpful to receive the electronic version of the screening results.
They also deliberate further that they (85%) preferred an automatic updating of results
which would be sent directly to their database. The least (70%) preferred method would be
a manual updating method utilising a Compact Disc (CD) with results burnt on it and sent
to the surgeries through the postal services. Incidentally, the GPs’ preferences were split
equally on supervised and unsupervised updating; in this case the results would be sent to
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
81
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
GPs similar to an email message and require supervision to accept the updating to the
woman’s record.
GP intervention
Three quarters 45 (76%) of the GPs suggested that interventions initiated by them have a
strong impact on increasing screening attendance. Nearly half 20 (49%) of the GPs who
responded chose (one to one type) opportunistic intervention as the most suitable form of
intervention. Just under a third (17) believed that making an intervention by post to the
non-attending women would be more suitable. 13 (23%) preferred a telephone-based
intervention. Of those who preferred opportunistic intervention, almost all 18 (98%)
consider that the intervention would not take more than ten minutes. While selecting the
preferred surgery staff for intervention, 45 (88%) respondents believed that nurses were
appropriate to conduct an intervention in the surgery and 35 (75%) considered the doctor
was more suited for intervention. In addition, 26 (60%) of the GPs indicated that
appointment booking assistants were preferred interveners. Surgery demography was not
significantly associated with the GPs who thought that there is strong impact through
intervention (Table 16)..
Additional information on interventions
Three quarters of the GPs 43 (73%) responded that they have the ability to intervene if
provided with the knowledge of non-attendance. Bivariate analyses show significant
association between the GPs who said that their ability to intervene can make a positive
impact in increasing uptake and the number of women population registered with them
(86% vs 61%, p=0.035) (Table 16). The majority of GPs 40 (80%) wanted recognition for the
efforts put in by the surgery for increasing the screening uptake. Out of the GPs who
believed that their efforts are to be recognised also regarded that appropriate awards 16
(26%) and proper appreciation 29 (48%) were recommended by them. Almost 39 65% of the
GPs thought that their efforts should be compensated through remuneration. Although the
difference did not reach statistical significance, surgeries with more than two partners were
more likely to accept that GPs interventions have impact to increase screening attendance
than those surgeries with only one physician (82% vs. 62%, p=0.082) (Table 16). The GPs
responding to this section also indicated that additional assistance is needed in the form of
patient directed leaflets on breast cancer in English 43 (72%), leaflets in other languages 39
(65%) and more training to GP staff on breast screening (58%) in order to make better
interventions for increasing breast screening attendance. Fifty percent 35 (50%) of GPs also
expected more assistance in the form of videos on breast screening and GP info-packs.
Questionnaire results
As part of a wider research, this survey was directed at finding the associations of various
attributes with screening results, predictions and GP interventions.
Many of the
respondents were female GPs; this is of some importance as this questionnaire was not
directed specifically to any gender. The practice manager to whom the questionnaire letter
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
82
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
was addressed might have expected female GPs would be more interested to answer the
questionnaire yet there were sizeable practices 22 (39%) with two or less physicians
practising. The national average of two or less partners in a surgery is around 20% (RCGP
2005). The number of female GPs in UK has increased from 23% to 35% (RCGP 2005); all
these factors might be the reasons for more female GP responses. The number of women
aged 50 to 70 registered with a GP was expected to be evenly distributed according to the
GPs who participated in the questionnaire piloting phase. Although the respondents had
indicated that the majority had more than 400 women registered with them, the data
collated for the same age group from the Royal College of General Practitioners (RCGP)
information sheet has indicated that, on average, the number of women registered with UK
GPs is approximately 200 (The Information Centre 2007, RCGP 2005). The figures collected
in this questionnaire, which is double the national average, is a clear indication that this
sample is not truly representative, yet there is a sizeable women population in the study
area who are eligible for breast screening and who are also registered with GPs. Increasing
consultation rates from 5 to 7 consultations per person year) from 1995 to 2006, and a rise
with age (Table 15) indicates that GPs do have the opportunity to address screening
attendance through opportunistic intervention (Figure 58).. Jepson et al. (2000) considered
that educational home visits to be one of the effective interventions. The NHS information
centre’s study indicates that the number of GP home visits halved (from 9% to 4%) (The
Information Centre 2007) reducing the opportunity to intervene in a non-surgery
environment. On the other hand telephone-based consultations have trebled (from 3% to
10%). At the same time provided additional possibilities for making opportunistic
intervention to increase screening attendance.
Figure 43 - Consultation rate per person year trend (1995-2006) for age group between 50 and 69
The study also reiterated that the GPs agree that the majority of the women who are aged
between 50 and 70 approach the GPs for almost all the listed conditions such as
muscular/skeletal disease, diabetes, arthritis and hypertension. Nearly 15% of the
respondents also indicated that this statistic was too difficult to collate or requires more
efforts to identify the numbers. Some of the GPs also reported that the women of the same
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
83
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
age also visit more often for minor illnesses such as infections, stress related problems and
upper respiratory tract infections, to specify a few. Earlier studies by Sin and Leger (1999)
indicated that clinical staff often makes house calls / visits to those women in the age group
50-70. It is imperative that the current research (reported in this paper) identifies the staff
making house calls in the primary healthcare team. More information pertaining to the
frequency of their house visits would be helpful to recognize the possibilities for making
interventions in a non-surgery type environment. Nearly half of the doctors make at least 6
or more visits (per month), 15% of nurses make 6 or more house calls and 11% of special
nurses make 6 or more home visits. Out of these, the doctors were the preferred choice of
clinical staff to make such kind of house call. The house calls are expected to be conducted
in a more relaxed environment and this could yield a good doctor-patient relationship
(Hinton et al. 2004).
This creates a unique opportunity for the doctor to have a personal rapport with their
patients and hence is more suitable for making an opportunistic intervention to address the
screening non-attendance issue (Sin and Leger 1999). Further analyses identified that the
GPs who responded positively about the impact of GP intervention can be predicted by the
clinical staff house visit calls attribute (R= 0.494 and χ2 =9.519, p<0.02). How ever
insignificant, one GP mentioned an interest in intervening but is handicapped through lack
of time, this shows that some GPs may have intent to intervene but are restricted due to lack
of resources. Predicting women’s breast screening attendance and sharing the knowledge is
a new concept. The response of the GPs regarding the usefulness of prior knowledge on
non-attendance (so that they can intervene) is an important issue. Most of the GPs
considered that prediction would be useful. This had strengthened the assumption that GPs
themselves believe that they accept the idea of receiving the a-priori knowledge of nonattendance. The attributes for predicting the GPs acceptance of prior knowledge is
dependent on those who also believe that GPs have the ability to make a positive impact
through intervention to increase breast screening uptake (R= 0.450 and χ2 =11.816, p<0.001).
An important aspect of this research is to identify at what point the AI-based prediction
should be run and sent to GPs. This should neither be too short - which would result in
limited time available for an opportunistic intervention to take place - nor too long to the
point of diluting the fervour of intervention. The respondents overwhelmingly agreed that
about four weeks in advance would be appropriate for initiating intervention. These four
weeks are well within the screening process schedule. A few (less than 10%) responded
positively for thirty or more days prior to the date of screening. A two-pronged strategy
recommended by this project (Baskaran et al. 2005) aims not only to gain advantage of
prediction knowledge but also the knowledge gained through the post-screening results on
non-attendance through appropriate intervention. The said intervention may happen over
time as it would be conducted after the completion of the screening cycle. All these factors
warranted further evidence on how the screening results are received by the GPs, their
updating frequency, mode, etc. At present, the screening centre mails the screening result
copy to the women and to their GPs. The screening results are in paper format and are filed
with the woman’s clinical record at the GP surgery. The respondents have unanimously
agreed that the screening result (in paper format) is very useful in providing better
healthcare to women. During the questionnaire piloting phase; there were some
apprehensions about the usefulness of the screening results and possibility of the results not
being handled appropriately.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
84
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Hence the GPs were also asked whether they obtained the screening results and most of
them concurred that they receive the results, and the same number of respondents
recognised that those results were always updated to the women’s records. Out of the
sample who responded, 3 (5%) answered “No” and 2 (3%) answered “Did not know” as to
whether the results were received, so there is still scope for improving the process of
screening results delivery. In the UK, primary care has been computerised and the majority
of surgeries are utilising computers to assist in their care delivery (Protti 2005). This was
also reflected in the responses to the question on the mode of surgery updating. Twelve
percent (12%) of the GPs suggested that the paper format results are scanned as files and
stored directly in their database, and half of the GPs responded that the results were
manually updated as text with ‘Read Code’ into the women’s records by either the GP
themselves or by the practice manager or nurse or, in some instances, by the data clerk.
Computers play a vital role in the surgeries and provide assistance for laboratories to send
the results directly to the GPs in electronic format (Protti 2005); most of the GPs (86%)
agreed that it will also be helpful to get the screening results in electronic format.
Other screening services utilise Compact Disc (CD) for transferring the results of screening
to the GPs. This format facilitated database updates; this not only saved time but also
avoided errors while updating, still the transfer to the computer system needed manual
involvement. 27 (45%) of the GPs did not answer this option. Supervised updating of
results can be accomplished when the results are being delivered such as an e-mail message.
Supervised updating mode response was identified as a predictor for GPs who agree that
prediction is useful for increasing attendance (R= 0.362 and χ2 =4.221, p<0.04). This question
was also not answered by as many as half of the GPs who responded. Almost 85% agreed
that the results, if sent directly to their database, would be the most preferred option. GP
intervention is the core issue of this questionnaire. The GPs should have faith that the
interventions made by them have strong impact on increasing screening attendance (Taylor
et al. 2003). This was reflected by the positive responses 45 (75%). Earlier research by
Richards et al (2001) suggested that opportunistic intervention was more economical and
effective (Bell et al. 1999). Forty seven percent 27 (47%) of the GPs concurred with the
opportunistic (one-to-one) type of intervention as preferred. To our surprise, the
traditionally preferred letter-based intervention (Turner et al. 1994, Majeed et al. 1997) for
screening was also preferred by 17 (30%) of the responding GP. One of the earlier studies
identified that even the appointment booking assistant, if trained appropriately, can play a
positive role in increasing screening attendance (Atri et al. 1997) and this idea has been
strengthened by the GPs who responded to our questionnaire: they preferred appointment
booking assistants (60%) as suitable intervening staff. Healthcare assistants were also
suggested by one of the GPs as a preferred person to conduct the intervention.
Multi-variate analyses of preferred intervention personnel (nurses and booking assistants)
attributes indicated with statistical significance (R= 0.507 and χ2 =10.388, p<0.006) that they
were positive predictors in identifying the GPs who felt that prediction was useful. The
predictive accuracy increased when the doctors were also included in the set of attributes
(R= 0.578 and χ2 =13.429, p<0.004); this increased the prediction’s statistical significance.
One of the GPs strongly responded that there were no resources available to allow
involvement in any kind of intervention. The knowledge of non-attendance has been
perceived to make a positive impact through intervention to increase breast screening
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
85
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
uptake (Sin and Leger 1999). Nearly 26% of the GPs either answered “No” or “Did not
know” for this query. The GPs who thought that their efforts for making intervention has to
be recognised have also preferred remuneration 39 (65%) as the most recommended form of
recognition followed by appreciation 29 (48%) and awards 16 (26%).
The qualitative study conducted during the piloting phase added proof to the findings of the
earlier study by Sin and Leger (1999); both had indicated that remuneration-based
recognition was the most preferred method. However, triangulation of the results to the
questionnaire respondents revealed a different view; appreciation and awards were
preferred by some GPs. These questions were aimed to find two decisive factors. 1)
Whether the GPs felt the need to be recognised for their efforts to make intervention? 2) Do
they feel that intervention was part of the general care delivery? The questionnaire
respondents clearly indicate that some form of recognition is vital for making the
intervention-based increase to screening uptake and GPs were of the opinion that they
cannot do the intervention as part of general care delivery due to the lack of resources.
Analysing the questionnaire, some GPs asked for additional resources (within their surgery)
to make an impact on screening attendance. The questionnaire response also supported this
aspect as about three quarters of the GPs (74%) recommended all the listed options in the
questionnaire. In order of preference, the GPs thought that they require more patientdirected leaflets on breast cancer in English and in other languages, training of GP staff, GP
info-packs and videos about breast screening. Even though most of the above mentioned
that assistance is being provided by the NHS, still there is scope either to increase the level
of assistance or to cover more surgeries with said assistance.
There were certain limitations to the current study. This questionnaire was aimed at small
geographically related GPs and had less than average responses; hence it might not
accurately reflect the entire GP population’s views in the UK. This was the reason that most
confidence intervals had wide ranges. Interventions were assumed to improve static
screening uptake figures. Moreover women are encouraged to make informed decisions
about attending a breast screening episode (Turner et al. 1994). Hence interventions alone
may not be the contributing factor for improving the static uptake figures. This research
relies heavily on the success of the proposed NHS-CfH’s IT project and the continued use of
computers in the primary care domain. Prediction through AI is not 100% accurate and
errors are expected in the form of false positives and false negatives. Some of the questions
had very low response rates which might warrant a more detailed study on the specifics.
The questionnaire has strongly suggested that there is a high probability that screening aged
women visit the surgery for various health related reasons and hence provide an
opportunity for the GPs to intervene. This study clearly suggests that GPs strongly believe
that they can play a vital role in increasing the screening uptake. Owing to the fact that, in
the UK, the primary care deliverers have been transformed into small business unit type of
organisations, this had a marked impact on their way of healthcare delivery. GPs believe
that all the time they have is to be spent in monetary oriented clinical services. This may
have increased the overall efficiency of the surgeries and their care delivery effectiveness,
but resulted in the non-remunerative and indirect primary care related services, such as
screening, taking a back seat. Some screening services such as cervical screening are offered
by the surgeries and they are compensated for the services rendered to the women by the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
86
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
NHS. Similar compensations are recommended to make the proposed breast screening
attendance based interventions a success.
The UK NHS has embarked on its most ambitious IT project and it is building the
infrastructure for providing automated Electronic Data Interchange (EDI) across its entire
clinical domain. Our study hypothesised that the utilisation of the infrastructure developed
by the NHS Connecting for Health (CfH) can be effectively used to deliver the predicted
screening attendance knowledge through Health Level 7 (HL7)-based messages (Baskaran et
al. 2004, 2006a, 2006b, Protti 2005). Continuous improvements to computer networks and IT
infrastructure are the norm of modern healthcare. This would not only be highly beneficial
for increasing the uptake but would also be economical. The GPs have provided this
questionnaire with an overwhelming positive response on their willingness to automate the
mode of message delivery.
Survey summary
All these factors and the response to the questionnaire have suggested the following. If such
kind of interventions (based on prediction or post-screening results) are to be made
effective, some form of compensation to cover economical overheads in surgeries is
essential, otherwise these initiatives are destined to fail. The same information was reflected
during the qualitative study conducted earlier. The responses did not provide a clear picture
on the preference of opportunistic intervention over the traditional telephone- or letterbased interventions. More focused studies in this area can shed light on these aspects.
Earlier studies on computer generated reminders for GP interventions had suggested an
increase in screening uptake (O’Connor et al. 1998, Shea et al. 1996). Hence an electronic
version of knowledge transfer to flag reminders is highly favoured by the GPs. The
questionnaire results (on the mode of knowledge sharing) concurred with the findings of the
earlier qualitative study on the same issue. This augmented well with the research strategy
adopted for this study; the triangulation of the results provided a benchmark which
supported the researcher’s beliefs on GP interventions (to increase screening attendance). All
the above inferences are very encouraging and provide the required evidence that GP
interventions are effective and can improve screening attendance further, it reveals that
computer initiated interventions would be economical and effective.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
87
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Conclusion and Future Work
This project set out to answer the research question and deliver the objectives mentioned in
chapter 1. This project not only confirms that breast screening attendance can be predicted
through an automated software solution but also can be leveraged to increase screening
attendance by employing emerging KM tools and techniques. This work has also proved
that such techniques can be implemented using current technologies available at the NHS’
disposal.
A prototype model (encompassing the JAABS algorithm) was designed and implemented on
Open Source technology as part of the project. This prototype model has become the first
iteration in preparing a complete software–based solution for addressing the breast
screening attendance. The project also validated the prototype through training and testing
with thirteen years of historical screening data from the Warwickshire, Solihull and
Coventry Breast Screening (WSCBS) unit. The first hypothesis of this project “that it is
possible to design and implement an Open Source-based prediction algorithm prototype for
breast screening attendance” has been tested and validated using the current datasets. The
results were then triangulated with the AI-ATT algorithm’s test results.
In addition, the proposed knowledge transfer was accomplished using appropriate
messaging standards (Health Level 7). Further, a detailed description on the design and
implementation of the new message suite (intended exclusively for sharing knowledge with
the primary care deliverers) was provided. The new Breast Screening Attendance Messaging
Protocol (BSAMP) that can be implemented on the National Care Records Service (NCRS)
being pursued by the NHS was depicted earlier. The second hypothesis “it is possible to
design a messaging platform for knowledge creation (through prediction) and sharing
model (knowledge/information exchange architecture) with available technology and
infrastructure” has also been tested and a workable model has been highlighted earlier.
Implementing the BSAMP on NCRS would require no additional investments. This lends
support to the research proposal’s ability to be integrated with minimal effort from
screening staff and GPs. Such initiatives would not only be economical but could be rapidly
deployed. A survey of the physicians for the WSCBS area confirmed that GP interventions
are feasible through electronic prompts and such interventions will undoubtedly improve
breast screening attendance.
The research also highlighted the advantages of applying both qualitative and quantitative
research methodology in combination. The synchronisation of these methodologies not only
supported each other but also provided a baseline for comparison using triangulation
techniques. Furthermore, the research approach adopted for this highlighted the importance
of conducting qualitative research. The qualitative research strategy also clarified the
research direction taken and offered support for further quantitative research at a later stage.
This work also presented the need for a balanced research methodology for maximising
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
88
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
deliverables. The questionnaire tested the last hypothesis “it is feasible for GPs to make
interventions to increase breast screening attendance through electronic prompt” and the
results were discussed.
Conceptual development
One of the original aims of the current research was to employ KM and its associated
technologies to increase the breast cancer screening attendance. A true KM-based project has
to address all the aspects of knowledge and its components. This section explains the
development of KM-related concepts so that an original contribution can be made to the
body of new knowledge. The following section has been structured based on the three core
components detailed earlier.
Concept development - Knowledge management
The first phase focussed on a critical literature review which highlighted the clear need and
scope for redefining the very essence of KM. This prompted a new examination of how
knowledge needs to be viewed. In particular, the differentiation of explicit knowledge and
information was of paramount importance when applying KM to the healthcare field. The
concept and efficacy of KM within healthcare is not always understood and the notion of
such critical terms as “information” and “explicit knowledge” are often (incorrectly)
regarded as being interchangeable. We elaborate later on this concept through an easy to
understand clinical example. The view proposed as first and second generation information
is a new concept and clearly differentiates the differences between these terms. Equipped
with this new understanding, even novices can clearly comprehend these fundamental
elements and avoid using then as misnomers. This ensures that knowledge and its
components will be properly addressed in both clinical and non-clinical domains.
Concept development - Artificial Intelligence (AI)
This new concept refers to the Artificial Intelligence algorithm which was developed for the
first time on an open source platform for the NHS breast screening programme. This
platform was pursued due to fiscal limitations and commercial licensing difficulties. Even
though applying an AI-based algorithm for breast screening prediction was not new
(Arochena, 2003), the proposed JAABS algorithm is different from its predecessor (AI-ATT)..
The primary difference is that the previous algorithm worked on a commercial software
platform and used a number of disjointed software tools for prediction. In addition, it
depended exclusively on the developer’s expertise to connect these tools and work as a
comprehensive prediction tool. The new JAABS algorithm is not only built on open source
technology but was completely automated, requiring no human involvement for completing
the breast screening attendance prediction.
In addition, this concept also identified a new attribute for improving the prediction
efficiency of the algorithm. This attribute was developed based on a novel approach which
employs postcodes (of the screening unit and screening women). These postcodes were
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
89
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
transposed to their corresponding latitude and longitudinal values. Using these values, the
distance between the two locations has been calculated and consolidated as a categorical
variable and input to the AI network for breast screening attendance prediction.
Concept development - Health Informatics
The last conceptual development relates to the health informatics aspect of this work. The
NHS has initiated the design of HL7-based messages for a complete gamut of messaging
domains. This message development process is undergoing a phased delivery and is yet to
cover the breast screening domain. The current research not only created a set of messages
using HL7 version 3 protocols but also proposed a new message design template using the
very tools HL7 UK recommends for developing them. This unique template corresponds to
knowledge sharing both at the pre- and post-screening stages. The concept of employing
HL7 messages for knowledge sharing to address breast screening attendance is the first of its
kind. This exercise has proved that developing HL7 messages, for even a limited scope, is
possible with currently available software tools.
Original contribution
In this project, several original contributions have been made to the body of knowledge
associated with healthcare. The novel concept of screening attendance prediction in an Open
Source environment was new to the breast screening domain. In addition, an innovative
prototype was constructed to investigate the validity of the idea proposed through
independent research and analyses. Eardley and Czerwinski (2007) in their paper had
highlighted the importance of KM in healthcare and how the NHS has embraced and
adopted these tools, systems, and strategies for many KM initiatives. This research draws its
strength from such KM tools and techniques. This work is also one such initiative
addressing the NHS’ breast screening attendance through efficient KM methodologies.
This research has yielded many ideas; in particular, a fundamentally new concept of
Knowledge Management (KM) and its constituents were identified. Such radical new
concepts are going to revolutionise how future knowledge specialists in health domain are
going to view and apply KM for better care delivery. A new, simple and ready to use KM
framework and model were proposed and this project used such a model as part of its
validation. Such new concepts and models are not limited to healthcare domain only. These
models can be easily adapted to any business domain with minimum amendments. This
research was funded by the NHS Cancer Screening Programmes. For the first time, this
project had compared the most common healthcare messaging standards employed in UK.
A comparison table, listing the differences between the Electronic Data Interchange For
Administration, Commerce and Transport (EDIFACT) and Health Level 7 (HL7) standards
is presented in an earlier section of this project.
As this project work was part of a larger research group’s activities, it extended the
previously completed doctoral research through continuous and focussed research. This
type of continued research not only adds value to the cancer care and breast screening; it
also identifies new ways to further the knowledge boundary through iterative, selfAll enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
90
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
sustaining, and focussed research cycles. In doing so, it also validated the previous research
with updated set of data thus gaining more credibility to both the initiatives (AI-ATT and
JAABS). The empirical work completed on KM, AI and healthcare messaging front has not
been experimented in such a combination before for increasing screening attendance. This
work ideally would be the first of its kind in the breast screening domain. This work would
attract further research for continuously improving the care delivery process related to
breast cancer at the national level.
From the context of multi-faceted research, this project is highly interdisciplinary in nature.
This work that had captured emerging technologies such as HI, AI and KM, creates a
synergy in these fields for addressing breast screening issues. All of these said new
contributions were found worthy of publishing in international conference proceedings and
journal publications. Such peer reviewed publications has certainly augmented the value
and added the quality to this project. A summary of some these original contributions are
listed in Table 17.
Competency
Evidence of an original
investigation
Competence in independent
research
•
•
Corresponding project work
prediction knowledge sharing
new predictor variable
•
KM in healthcare
•
neural network prototype implementation
•
questionnaire-based survey on GP
interventions
•
HL7 v3.0 messaging in breast screening
A thorough understanding of
the appropriate techniques
Qualitative and quantitative research techniques
were employed for neural net testing and GP
knowledge sharing
Ability to make critical use of
published work and source
materials
Valuable references such as Richards et al.,
Bankhead et al., Kalra’s PhD project on EHR,
Dwivedi’s project work on KM framework are to
name a few, for a more complete list of source
materials, refer to the reference list
Appreciation of the relationship
of the special theme to the
wider field of knowledge
Worthy, in part, of publication
Setting down a major piece of
new information
•
core concepts of knowledge
•
KM framework and models in healthcare can
be easily applied in any domain
•
seven conference proceedings
•
two journal papers were published
•
two more journal papers are work in progress
•
refer to the list of publications for more
details
•
applying KM in real-time healthcare projects
•
comparison of healthcare messaging
standards
•
breast screening attendance predictionbased messaging are addressed for the first time
Continuing a previously original
piece of work
Arochena’s pioneering empirical work on breast
screening attendance prediction was continued
to prototype level
Showing originality in testing
somebody else’s idea
Arochena’s neural network ROC characteristics
was tested on new dataset and extended to new
episodes
Carrying out empirical work that Empirical study on GP’s views, attitudes and
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
91
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
has not been done before
beliefs on screening attendance prediction based
interventions
Being cross disciplinary and
using different methods
This project work is a true cross-disciplinary
work spanning areas from KM, AI, health
informatics, healthcare, ICT
Adding knowledge in a way not
done before
This project is a seminal work on predictionbased screening attendance improvement, using
available technologies through the appreciation
of humanistic nature of knowledge and its
management
Table 17 - Project work's originality summary
Objectives accomplished
This section details the accomplished objectives. It compares them to the list (set in chapter
1) of intended objectives.
•
•
•
•
As part of the research methodology, semi structured interviews were conducted
(through open-ended questions) which assisted in identifying the appropriate
research direction and focus.
The BSP call/recall process was thoroughly
investigated as was the viewpoint of GPs (as part of the interview-based pilot study);
it was crucial to understand the opinions, beliefs and expectations of these
“knowledge receivers” to employ effective interventions as a strategic tool for
increasing screening attendance.
The literature review investigated the suitability of Electronic Data Interchange For
Administration, Commerce and Transport (EDIFACT) and Health Level 7 (HL7)
standards. As no comprehensive comparison of these standards were discovered
during the literature survey, the project incorporated a comparison table. A
comprehensive report which included design methodologies and implementation
details for both standards was also included in the project report. This report
highlighted the appropriateness of HL7 version 3 for this knowledge exchange.
The design and analysis of a questionnaire to evaluate the GPs’ role in reducing BSP
non-attendance was accomplished. This questionnaire not only identified the GPs’
opinions, beliefs and attitudes for reducing non-attendance through intervention but
also collated the required information (for the correct timeframe) to share the
predicted knowledge. The most important finding of the questionnaire was that not
all GPs were expecting to receive financial compensation for their efforts to address
screening non-attendance. A sizeable number of GPs reported that they would
initiate interventions for any form of appreciation.
The project identified a messaging platform and designed a viable architecture for
integrating the Java-based Attendance prediction by Artificial Intelligence for Breast
Screening (JAABS) algorithm. This algorithm was not only tested with a dummy
dataset but also used real-time data collected from 1995 to 2008. Complete
automation was incorporated within the proposed protocol right from the stage of
raw data processing to the final prediction list. As a continuous improvement
initiative, a new predictor attribute was also added to the existing predictor set. In
addition, the quantitative analysis validated (through triangulation methodology)
JAABS and AI-ATT to Logistic Regression.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
92
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
A complete Knowledge Management-based protocol (BSAMP) was developed ready
for knowledge sharing with healthcare stakeholders via the NHS’ National Care
Records Services (NCRS) infrastructure. The BSAMP protocol included the JAABS
algorithm as the knowledge creation component. This project work was not limited
to the design of the architecture but also created real-time message templates that can
be instantiated to deliver the predicted knowledge. These templates covered both the
pre and post-screening knowledge sharing stages. These templates were created
using the HL7 version 3 message developing tools suggested by HL7 UK and NHS
CfH.
Project deliverables accomplishment list
•
•
•
•
•
•
•
•
•
evaluation report on the suitability of HL7 v3.0 standard
comparison of HL7 v3 vs EDIFACT
BSAMP message design and instantiation
architecture for proposed knowledge sharing through messaging
implementation details of JAABS algorithm
design for prototype (JAABS algorithm)
prototype execution and testing
questionnaire survey results and recommendations
reports, publications and presentations (refer to list of publications)
Adopting the proposed protocol
Screening
Year
Number of
women
invited
Screening
nonattenders
Probable
number of
cancers not
detected
Prediction of nonattenders using
JAABS @42% NPV
rate (till 6th
episode)
Approximate number of
women’s life saved,
even If 25% of the
interventions were
effective
1996-97
1,558,995
388,190
1,776
728
182
1997-98
1,668,476
408,777
2,635
1,080
270
1998-99
1,669,727
409,084
2,580
1,058
264
1999-00
1,811,541
424,951
3,110
1,275
318
2000-01
1,815,610
453,903
3,286
1,347
336
2001-02
1,752,526
429,369
3,237
1,327
331
2002-03
1,873,470
473,988
3,539
1,451
362
2003-04
1,998,989
495,750
4,284
1,756
439
Table 18 - Probable number of life saved by JAABS algorithm
The objective of this project was to identify the challenges which are being faced by the UK
NHS’ national breast screening programme and find approaches to alleviate these
impediments and eventually reduce mortality due to breast cancer. While conducting the
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
93
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
literature review, the static attendance was identified as a persistent challenge and
addressing it through KM was the driving force behind this research work. Table 18
provides a hypothetical scenario with current data to highlight the screening nonattendance, probable number of cancers not detected and approximate number of lives that
can eventually be saved through the proposed initiative. This table provides an insight into
the algorithm’s efficacy in saving human lives. When the proposed JAABS algorithm is
deployed as envisaged in this research, it can considerably decrease breast cancer mortality.
Based on the JAABS algorithm’s negative prediction value (for the first to sixth episode) the
number of non-attendees correctly predicted a priori to screening date would be at least 42
women for every 100 screening non-attendees. When such knowledge is shared with the
GPs (with whom the women are registered with) can initiate interventions. Such
interventions can educate the non-attending women and clarify their attitudes and beliefs
(Bekker et al. 1999). The expected outcome is that the women commits to a positive informed
decision, which would culminate in attending the screening appointment.
A 25% success in GP interventions will result in saving more than 350 women’s lives per
year. Even if one women’s life can be saved by our approach, this approach can be deemed a
success. The new bespoke software prototype, incorporating JAABS algorithm can be easily
converted and integrated in to the NBSS software. The messaging part has not been
prototyped, as it has been envisaged to utilise the NCRS architecture operating on SPINE
and DTS services. These services are already being field-tested at cluster level and once the
national network is established would form the backbone of all NHS’ future clinical
messaging across the country including the proposed protocol (BSAMP).. The current work
has already designed HL7 v3.0 messages for delivering the knowledge package as part of
the new protocol (BSAMP). This message design has been developed on the same modus
operandi followed by NHS CfH for easy integration. The ready to use message would ensure
rapid deployment of the BSAMP at the national level with little overheads and
modifications. As part of the current work, a questionnaire-based survey conducted with the
GPs has identified the need of additional resources and recognition for making such future
endeavours a complete success. Any future BSAMP-based initiatives for addressing the
screening attendance should undeniably approach the issues from the KM perspective for
ensuring positive results in improving breast screening attendance.
Challenges
Any research at the doctoral level has to work on the periphery of the corresponding
domain’s knowledge boundaries. This creates numerous challenges to operate with limited
resources. The requirement of continuous innovation at the highest level is fundamental for
research work. Such exercises will certainly enable in developing new ideas, strategies and
models. This section describes the challenges encountered in this research and how these
were dealt with and solved. A number of intriguing challenges were encountered and
solved during the current research work. This research provided a unique opportunity to
better understand the nuances of applying KM in healthcare. It assisted in realising that
there are numerous activities which business domains expedite as part of their routine
activities even without knowing their importance and its relationship to effective KM. The
“people” concept in KM is paramount and is the driver for any successful KM project.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
94
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
GPs’ perceptions on primary care and their apprehensions on how their care services
delivery have been transformed into small business domains and their impacts on
healthcare were highlighted. This had placed in a lot of undue pressure that can result in
counter productivity. GPs are the backbone of health services and are in direct contact with
the public and their efforts are to be valued. Due recognition, even simple appreciation, can
be leveraged as great motivators. These fundamental factors are all KM driven concepts of
people being the core of any knowledge related activities and services. During the execution
of this research work, it was found that convincing the stakeholders with regards to the
importance of KM in healthcare was challenging. The GPs were reluctant to spend their
resources on non-remunerative services i.e. addressing screening attendance. Future
screening initiatives on screening attendance have to address these issues at a core level;
failing to do so will force such initiatives to fail from achieving their true potential.
The current study relied on cancer data related to screening women’s historical episodes and
demographic details for training the JAABS algorithm. This necessitated ethical approval for
the whole research. This was achieved first by applying and obtaining the consent from the
university’s ethics committee. Following this, the NHS Local Research Ethics Committee
(LREC) was approached for getting approval. Even though the approval process was
lengthy, and time-consuming, the whole exercise gave an opportunity to better understand
the ethical issues related to clinical and non-clinical research.
The LREC approval meeting not only assisted in explaining the current research’s impact on
screening attendance to the committee members, but was seen as an opportunity for
acquiring new perspective and out-of-the–box ideas from connoisseurs having diverse
expertise. The WSCBS unit had insisted to keep the dataset which contained the women’s
postcode within the screening premises to ensure privacy, data protection, and security. To
address this issue the prototype’s data pre-processing module was made as an individual
entity that can consume the flat file generated by the screening unit’s system. This data preprocessing module processed the input data and the postal code was replaced with two
attributes namely the ‘postannum’ value, the “Townsend” deprivation value. The screening
distance was calculated and a new variable containing only the distance-based categorical
value was generated. Finally, this module gave an in process file that can be ported to the
laboratory environment for further processing such as training, validating and testing the
JAABS algorithm.
Designing and implementing the JAABS algorithm within an Open Source environment was
challenging. This can mostly be attributed to the fact that most of such earlier work in the
1990s was carried out in C and C++ environment. Availability of efficient commercial
software platforms have reduced the need for development of bespoke software, such as the
one which was created for JAABS algorithm. This has not only created a gap in the
availability of experts in this area but also has not given opportunities for creating such
software solutions at research and development level. Commercial software platforms have
their own inherent advantages such as ready made, time tested applications, better
maintenance, support and a large choice of ready made algorithms that allows a mix-andmatch strategy for exploratory type of work. Simultaneously such platforms also come with
weaknesses, such as not being flexible; users need special training, lack of automation and
non-availability of capturing repetitive work with less or no human involvement.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
95
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
The Java environment being relatively new has less support for neural networks and hence
had few prior initiatives in the said domain. One major handicap of Java was the memory
recovery. The automated garbage collection made it difficult to release unused memory.
Neural networks require massive computing resources. This fundamental requirement puts
undue load on memory allocation and management, on the operating system and hardware
resources. This challenge was partly overcome by using basic data structures available in
Java and partly by creating in-process flat files for speeding-up the processing to avoid
massive data to be stored on the Random Access Memory (RAM).. The next challenging
aspect of this research was the messaging protocol. This research work clearly identified that
any new clinical messaging has to be focussed on futuristic needs of healthcare. Hence, the
BSAMP protocol was planned to be based on HL7 v3.0 standard. The challenges can be
attributed to very little earlier work done on this version and all the work being currently
accomplished within the NHS UK were within a private environment and not publicly
accessible; this resulted in immense challenges.
Such challenges slowed down the process of designing the HL7 v3.0-based messages for the
BSAMP. This was partly overcome by attending special training sessions organised by the
HL7 UK and its affiliated organisations. Further, the HL7 parent organisation has provided
free tools and guides for HL7 members and new message developers. In addition to this, the
support provided by HL7 UK and the NHS CfH was commendable; their assistance greatly
speeded up the message design and implementation for the current work. The above
section specifically summarised some of the challenges met and how they were resolved. In
addition to this, continuous changes in the NHS and process level enhancements
implementation within the breast screening domain necessitated frequent updating of the
strategy. Many micro issues that needed careful attention were duly incorporated by fine
tuning the research focus so that the research objectives were achieved.
Future work
The project work being the first of its kind was limited by resources; hence it had focussed
on core issues and culminated in to a series of steps for the future research activities. The
following section enumerates the probable future research direction that can be pursued to
increase the breast screening efficiency and further the knowledge boundary.
1.
The current work had concentrated on just one care deliverer i.e. GPs, for
making interventions. This research work can easily be extended to other care
deliverers such as specialist doctors for diabetes, cardiology, dentistry and
care deliverers such as optometrist, pharmacist, orthopaedist, and
physiotherapist. They are just to name a few who can make multi-pronged
interventions. Such interventions can add value by increasing the probability
of non-attendees to make an informed decision, thereby increasing screening
attendance.
2.
The algorithm is modelled in Open Source environment which employs
object oriented concepts. With little modifications, this can easily be extended
to predict other screening programmes actively pursued by the NHS National
Screening Programmes (bowel cancer, cervical cancer, diabetic retinopathy,
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
96
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
colon cancer, and prostate cancers). This may need exploratory research to
identify predictor variables for their respective domains.
3.
An immediate extension to the current work will be to identify a region and
link the prediction results using the message design template for messaging
through a web service-based email service to send messages to the GPs and
update their database. A randomised controlled trial on at least three groups
of GPs, one with the electronic messaging-based database updating, one
group with the traditional paper-based prediction results transfer and one not
making any prediction-based intervention. This study will validate the
effectiveness with the electronic transfer of prediction knowledge when
compared to traditional paper-based knowledge transfer.
4.
A probable way to increase the efficiency of the JAABS algorithm is to
identify more predictors. Ethnicity, educational background, women’s
economical dependency, other healthcare related attendance etc can be
included as predictors. This might need fresh data collection and collation for
the screening women population. In addition new prediction models such as
Support Vector Machines, genetic algorithm-based prediction techniques can
be used as enhancements to the neural network. This kind of hybrid
techniques will improve the current JAABS algorithm‘s Negative Predictive
Values (NPV).
Concluding observations
The current project has not only achieved all the planned objectives but also improved upon
them and added value by developing new concepts. The project work additionally confirms
that a research methodology can be an effective combination of both qualitative and
quantitative research paradigms. A hybrid research methodology, which incorporates both
paradigms, should also utilise triangulation as validating mechanism to achieve more
credibility. This report sets out to highlight the conceptual developments that were part of
the project and also clearly recognizes the distinguishing features with the earlier work
(Arochena, 2003). In particular, the importance of the health informatics and KM
components (which were not part of the earlier work) were added as new concepts to the
existing knowledge domain.
Even though the JAABS algorithm was an adaptation of the earlier work (AI-ATT), this
project focused on increasing the predicting efficiency and proved that such an algorithm
can be incorporated within an open source framework with complete automation. The
algorithm proposed in this project work is only a first step towards an automated prediction
mechanism that would be used in a day-to-day functioning of the Warwickshire, Solihull and
Coventry Breast Screening Unit. The ultimate goal is to extend the utilisation of the prediction
technique to all the screening units in UK. An important improvement that has been
implemented in the JAABS algorithm is to find more predictors for increasing its efficacy.
One such indicator can be the women’s ethnicity. This attribute is difficult to collect for all
the screening women population, but new research conducted in identifying the ethnicity by
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
97
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
the women’s name and other direct collation of information from the GPs and PCTs are also
viable alternatives (Szczepura 2005).
The research was also the first of its kind, enumerating simple knowledge-based models and
their implementation in preventive care. More than thirteen years of screening data was
utilised to train the algorithm for the first time. The current work also proved that complex
algorithms such as JAABS can be implemented in Open Source-based technologies. The
research provides a fertile ground for fostering new research to continuously improve care
services, specifically cancer surveillance. Clinical messaging is driven by the explosion of
information and communication technologies. Electronic health records are fundamental to
provide the right care, at the right place, at the right time and to the right person,
irrespective of time and place where the care delivery is required. Even though HL7 v3.0 is
relatively new and is not yet fully matured, further exposure and clinical implementation at
an international level will provide enough opportunities for the standard to evolve and
stabilise.
As part of this project, a qualitative analysis on health informatics and, in particular,
electronic patient records was seen as an added deliverable. The discussion focussed on
issues that are to be properly identified and resolved for effective electronic patient record.
Special mention of academie’s involvement in developing new talents and the need for
knowledge sharing at international level on health informatics issues were the focus of this
work. All these issues have to be actively pursued for making global electronic patient
record a reality. This project had revealed the effective use of primary care deliverers (GPs)
for initiating the intervention for increasing uptake. A suggestion to extend this messaging
framework to all the tertiary healthcare deliverers as interveners has to be validated and
tested. The intricacies involving a whole gamut of healthcare deliverers may throw up fresh
challenges and this can be tackled only by dedicated and well directed research in the said
field of study.
In an earlier section, it was highlighted how primary care providers can be crucial in making
preventive care a total success or failure. From the KM perspective, GPs are to be viewed as
knowledge workers who are strategically positioned to make a huge impact on the
population, thereby increasing the efficacy of such preventive healthcare services. Even
inexpensive approaches such as appropriate recognition can yield positive results. One such
approach of employing recognition (in the form of public appreciation of GPs efforts specific to the value added services provided - such as trophies, commendations and
awards) were proposed by the GPs. This work also proved that clinical messaging can be
employed for efficient knowledge sharing. The research work covered in this project has
indisputably underlined the importance of KM and seamless knowledge transfer across
clinical domains. It further described how it can be successfully employed at empirical
research. In addition, this work draws attention to the humanistic values that are to be
addressed for a successful knowledge based research work. It further described how people
(GPs), process (BSAMP) and technology (messaging, HL7, SOA, AI, XML) can work in
unison on a healthcare KM project.
Finally, the research had proposed the use of HL7 v3 based messages for EDI and
knowledge exchange. The HL7 v3 messaging has to be adapted for all future screening unit
message exchanges. Further projects in healthcare domain have to appreciate the leveraging
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
98
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
abilities gained through appropriate use of KM. It should be mandatory for all research
projects related to healthcare domain to focus on KM and its technologies to take healthcare
delivery into the twenty second century.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
99
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
References
Ackoff, R.L. (1989) ‘From data to wisdom.’ Journal of Applies Systems Analysis 16, 3-9
Adams, J., Ryan, V. and White, M. (2004) 'How accurate are townsend deprivation Scores as predictors of self-reported health?
A comparison with individual level data.' Journal of Public Health 27, (1) 101-106
Advisory Committee on Health Infostructure (1999) Paths to better health. Ministry of Public Works and Government Services,
Ottawa [online] Available from <http://www.hc-sc.gc.ca/hcs-sss/alt_formats/iacb-dgiac/pdf/pubs/1999-paths-voiesfin/1999-paths-voies-fin_e.pdf> [28 January 2008]
American
Cancer
Society
Inc.
(2005)
Cancer
eference
information.
[online]
Available
<http://www.cancer.org/docroot/CRI/content/CRI_2_4 _1X_What_is_breast _cancer_5.asp> [10 August 2005]
from
Ana, N. (2004) 'Achieving a healthy KM assessment.' Inside knowledge [online] 7, (10) Available from
<http://www.ikmagazine.com/xq/asp/sid.f519433d-2d2c-4a63-bfb1-d3496f0aaa7c/articleid.d7cbcffd-ff0f-4198-bd8417946ee17853 /etitle.achieving_a_healthy_km_assessment_nhs_modernisation_agency/qx/display.htm> [20 August 2004]
Anderson, J.A. (1995) An introduction to neural networks. Cambridge, MA: MIT Press
Anderson, R.J., Richards, B. (1996) Patient confidentiality at risk from NHS-wide networking, in current perspectives in healthcare
computing conference. Harrogate: BJhealthcareLimited, 687-92
Arbib, M.A. (1995) Handbook of brain theory and neural networks. Cambridge: MIT press
Argyris, C. (1992) On organisational learning. Oxford: Blackwell
Armoni, A. (2002) Effective healthcare information systems. New York: Idea Group Inc
Arochena, H.E. (2003) Modelling and prediction of parameters affecting attendance to the NHS breast cancer screening programme.
Ph.D. Project, Coventry: Coventry University (BL: DXN064476)
Arochena, H.E., Naguib, R.N.G., Mashhoudy, H., Wheaton, M.E. and Wallis, M.G. (2001a) 'Factors associated and predictors of
attendance to the UK breast screening programme.' Proceedings of the 23rd Annual International Conference of the IEEE Engineering
in Medicine and Biology Society
Arochena, H.E., Naguib, R.N.G., Mashhoudy, H., Wheaton, M.E. and Wallis, M.G. (2001b) 'Screening variation: its impact on
uptake and coverage for the breast screening programme.' European Journal of Cancer 37, (2) 110
Arochena, H.E., Naguib, R.N.G., Todman, A., Wheaton, M.E. and Wallis, M.G. (2002a) 'An algorithm for the prediction of
attendance to a breast screening unit in the UK.' Proceedings of the 2nd Joint Conference of the IEEE Engineering in Medicine and
Biology Society and the Biomedical Engineering Society 76-77
Arochena, H.E., Naguib, R.N.G., Todman, A., Wheaton, M.E. and Wallis, M.G. (2003) 'Blind study for the prediction of
attendance to a UK breast cancer screening unit.' Proceedings of the 4TH IEEE Information Technology Applications in Biomedicine
Conference
Arochena, H.E., Todman, A., Naguib, R.N.G., Wheaton, M.E. and Wallis, M.G. (2002b) 'Discussion of an algorithm for the
prediction of attendance at first invitation from a UK NHS breast screening unit.' Proceedings of the IEEE EMBS UK & RI
Postgraduate Conference in Biomedical Engineering and Medical Physics, IEEE EMB Student Society 17
Arochena, H.E., Whitford, A., Naguib, R.N.G., Mashhoudy, H., Wheaton, M.E. and Wallis, M.G. (2000a) 'Analysis of attendance
and screening variation in a UK breast cancer screening unit.' Proceedings 2000 IEEE EMBS International Conference on
Information Technology Applications in Biomedicine ITAB 2000, 332-337
Arochena, H.E., Whitford, A., Naguib, R.N.G., Wheaton, M.E. and Wallis, M.G. (2000b) 'Descriptive analysis of parameters
affecting a UK breast cancer screening unit.' Proceedings of the World Congress on Medical Physics and Biomedical Engineering
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
100
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Askevold, A., (2006) 'Secure healthcare messaging using open standards.' EuroPACS conference [online] Available from
<http://www.europacs.
net/Extended%20abstracts/Security%20and %20legal%20aspect/Annebeth% 20Askevold%20%20Secure%20healthcare%20 messaging%20using %20open %20standards.pdf> [10 March 2007]
Astion, M.L. and Wilding, P. (1992) 'Application of neural networks to the interpretation of laboratory data in cancer diagnosis
clinical chemistry.' Intl. Journal of Molecular Diagnostics and Laboratory Medicine 38, 34-38
Asuman, D., Gokce, B.L., Yildiray, K., Seda, U., Sam, H. et al. (2006) ‘Exploiting ebXML registry semantic constructs for
handling archetype metadata in healthcare informatics.’, Int. Journal of metadata, semantics and ontologies 1, (1) 21-36
Atri, J., Falshaw, M., Gregg, R., Robson, J., Omar, R.Z. and Dixon, S. (1997) ‘Improving uptake of breast screening in
multiethnic populations: a randomised controlled trial using practice reception staff to contact non-attenders’, BMJ 315, 1356-9
Augier, M., Shariq, S.Z. and Vendel, M.T. (2001) 'Understanding context: its emergence, transformation and role in tacit
knowledge sharing.' Journal of Knowledge Management 5, (2) 125-136, MCB University press
Bali, R.K. and Wickramasinghe, N. (2006) Managing knowledge: the NHS and beyond, knowledge management specialist library
[online] Available from http://www.library.nhs.uk/KnowledgeManagement/ViewResource.aspx?resID=187244 [28 January
2008]
Bali, R.K. and Wickramasinghe, N. (2008) 'Achieving successful EPR implementation with the penta-stage model.' International
Journal of Healthcare Technology and Management 9, (1) 97-105
Bankhead, C., Austoker, J., Sharp, D., Peters, T., Richards, S.H., Hobbs, R., Brown, J., Wilson, S., Roberts, L., Redman, V.,
Tydeman, C., Formby, J. (2001) 'A practice based randomised controlled trial of two simple interventions aimed to increase
uptake of breast cancer screening.' Journal of Medical Screening 8, (2) 91-98.
Baskaran V, Bali RK, Shah B and Guergachi A, (2009 a) "Knowledge management as a holistic tool for superior project
management”, International Journal of Innovation and Learning by InderScience (forthcoming)
Baskaran V, Bali RK, Arochena H, Naguib RNG, Wheaton M, Wallis M , Benson T and Wickramasinghe N (2009 b) "A
knowledge-based primary care approach to increase breast screening attendance", Int. J. of Biomedical Engineering and
Technology, 2(2) 172-88
Baskaran, V., Bali, R.K., Arochena, H., Naguib, R.N.G., Wheaton, M. and Wallis, M. (2006a) 'Improving uptake of a breast
screening programme: a knowledge management approach for opportunistic intervention', In Bos L et al (Ed) Medical and Care
Compunetics 3, IOS:Netherlands 191-7
Baskaran, V., Bali, R.K., Arochena, H., Naguib, R.N.G., Dwivedi, A.N. and Nassar, N.S. (2004) 'Towards total knowledge
management for healthcare: clinical and organizational considerations.' Proc of the IEEE-EMBC 26th Ann Int Conf of the IEEE Eng
in Med and Biol Soc (EMBS) 3163-3166
Baskaran, V., Bali, R.K., Arochena, H., Naguib, R.N.G., Wallis, M. and Wheaton, M. (2006b) ‘Knowledge creation using artificial
intelligence: a twin approach to improve breast screening attendance’, Proc of the 2006 IEEE Eng in Med and Biol 28th Ann Conf,
New York City
Baskaran, V., Bali, R.K., Naguib, R.N.G., Arochena, H. (2005) ‘A knowledge management approach to increase uptake in a
breast screening programme’, Second Humanoid, Nanotechnology, Information Technology, Communication and Control, Environment
and Management (HNICEM) International Conference, The Institute of Electrical and Electronics Engineers Inc. (IEEE) – Philippine
Section, March 17-20, HYATT Regency Hotel, Manila, PhilippinesHNICEM, IEEE, Philippines
Baumard, P. (1999) Tacit knowledge in organizations. Sage: London
Becker, M.Y. (2007) 'Information governance in NHS’s NPfIT: A case for policy specification.' International Journal of Medical
Informatics 76, 432–7.
Beeler, G.W. (1998) 'HL7 Version 3- an object-oriented methodology for collaborative standards development.' International
Journal of Medical Informatics 48, 151-161
Bekker, H., Morrison, L. and Marteau, T.M. (1999) 'Breast Screening: GPs’ beliefs, attitudes and practices.' Family Practice 16, 6065
Bell, D. (1999) ‘The axial age of technology foreword: 1999, in the coming of the post-industrial society’. New York: Basic Books
Bell, T.S., Branston, L.K., Newcombe, R.G.and Barton, G.R. (1999) 'Interventions to improve uptake of breast screening in inner
city cardiff general practices with ethnic minority lists.' Ethnicity & Health 4, (4) 277-84
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
101
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Bellenger, G. (2002) Emerging perspectives- systems thinking knowledge
<http://www.systems-thinking.org/kmgmt/kmgmt.htm> [10 March 2005]
management.
[online]
Available
from
Bennett, R.L., Blanks, R.G., Patnick, J., Moss, S.M. (2007) 'Results from the UK NHS breast screening programme 2000-05.' J Med
Screen 14, 200-4
Benson, T. (2002), ‘Why general practitioners use computers and hospital doctors do not- part 2: scalability.’ BMJ 325, 1090-3
Berdayes, V.W., Murphy, J.W. (2000) Computers, human interaction, and organizations: critical issues. Praeger/Greenwood
Berg, B. (2001) Qualitative research methods for the social sciences. Boston: Allyn and Bacon, 304
Berger, P.L. and Luckmann, T. (1967) ‘The social construction of reality: a treatise in the sociology of knowledge’. New York: Garden
City Doubleday
Bigus, J.P. and Bigus, J. (2001) ‘Constructing intelligent agents using java: professional developer's guide series’. Wiley Publishing
Binnie, L.M., Williams, J.M. (2002) ‘Intuitive psychological, physical and biological knowledge in typically developing
preschoolers, children with autism and children with Down's syndrome.’ British Journal of Developmental Psychology 20, (3) 34359
Bjurstam, N., Bjorneld, L., Duffy, S.W., Smith, T.C., Cahlin, E., Eriksson, O., et al. (1997) 'The gothenburg breast screening trial:
first results on mortality, incidence, and mode of detection for women ages 39–49 years at randomization.' Cancer 80, (11) 20919
Blacker, B.K. (1991) 'The development of an EDI system.' Computing &Control Engineering journal 231-7
Blackler, F. (1995) ‘Knowledge, knowledge work and organizations: an overview and interpretation.’ Organization Studies 16, (6)
1021-46
Blanks, R.G., Moss, S.M., McGahan, C.E., Quinn, M.J. and Babb, P.J. (2000) 'Effect of NHS breast screening programme on
mortality from breast cancer in England and Wales, 1990-8: comparison of observed with predicted mortality.’ BMJ 321, (7262)
665–9
Bojan, N. and Martin, B. (1995) 'Application of artificial neural networks for childhood obesity predicition.' 2nd New Zealand
Two-Stream International Conference on Artificial Neural Networks and Expert Systems (ANNES '95), 377
Bostick, R.M., Sprafka, J.M., Viring, B.A., Potter, J.D. (1994) 'Predictors of cancer prevention attitudes and participation in
cancer screening examinations.' Journal of Preventive Medicine 23, 816-826
Bradley, P., Bennett, K. P. and Demiriz, A.(2000) Constrained k-means clustering- Technical report., MSR-TR-2000-65, Microsoft
Research [online] Available from <http://research.microsoft.com/research/pubs/view.aspx? msr_tr_id=MSR-TR-2000-65> [10
March 2005]
Breakthrough
Breast
Cancer
(2006)
‘Corporate
social
responsibility.
[online]
Available
from
<http://www.breakthrough.org.uk/you_can_help/companies_and_ employees/whats_it_all_about/corporate_social.html>
[14 March 2006]
Breen, M. (1997) ‘Information does not equal knowledge: theorizing the political economy of virtuality.’ Journal of ComputerMediated Communication [online] 3, (3), Available from <http ://www.ascusc.org/j cmc/vol3/issue3 /breen.html> [10 March
2004]
Breen, R. (2006) 'Work based learning and the national health service university.' London: Kogan Page
Breen, R. (2007) 'An approach to mobilising knowledge within a new organisation.' NHS Institute for Innovation and Improvement
[online] Available from <http://www.library.nhs.uk/knowledgemanagement/ ViewResource.aspx? resID=235299> [23
September 2007]
Brennan, S. (2005) The NHS IT project: the biggest computer programme in the world . . . Ever!' UK: Radcliffe Publishing
Bresnen, M., Edelman, L., Newell, S., Scarborough, H. and Swan, J. (2003) ‘Social practices and the management of knowledge
in project environments.’ International Journal of Project Management 21, 157-66
Brown, J.S. and Duguid, P. (2001) ‘Knowledge and organization: a social–practice perspective.’ Organization Science 12, (2) 198213
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
102
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Brunton, M. and Thomas, D.R. (2002) 'Privacy or life: how do women find out about screening mammography services?' Journal
of the New Zealand Medical Association, 115, 1161. [online] Available from <http://www.nzma.org.nz /journal/115-1161/168/>
[10 March 2005]
Bryant, H. (1996) 'How should we interpret noncompliance with screening mammography?' Canadian Medical Association
Journal 154, 1353-1355
Bums, S. (2003) ‘Knowledge management in healthcare: succeeding in spite of technology’, Sbusiness journal, AFSM
International [online] 32-36. Available from <http://www.hmc-benchmarks.com/about/SBums_ ASFMI.pdf> [10 March 2004]
Burton, G. (1997) ‘Alternative medicine’. Washington, U.S.A.: Future Medicine Publishing
Calle, E.E., Flanders, W.D., Thun, M.J., Martin, L.M. (1993) 'Demographic predictors of mammography and pap smear
screening in US women.' American Journal of Public Health 83, (1) 53-60
Canada Health Infoway (2006) ‘EHRS blueprint, an interoperable ehr framework’. [online]
<http://knowledge.infoway-inforoute.ca/en/knowledge-centre/ehrs-blueprintv2.aspx> [28 January 2008]
Available
from
Canada Health Infoway, (2007a) ‘2015: Advancing Canada's next generation of healthcare’. [online] Available from
<http://www.infoway-inforoute.ca/en
/pdf/Vision_2015_Advancing_Canadas_next_generation_of_healthcare.pdf>
[28
January 2008]
Canada Health Infoway (2007b) ‘Corporate business plan 2007-08, electronic health records: transforming health care, improving lives’
[online] Available from <http://www.infoway-inforoute.ca/Admin/Upload/Dev/Document/ Business%20Plan_200708_EN.pdf> [28 January 2008]
Canada Health Infoway (2008) ‘Who we are?’ [online] Available from <http://www.infoway-inforoute.ca/en/WhoWeAre
/Overview.aspx> [28 January 2008]
Cancer
Research
UK
(2004)
Breast
cancer
factsheet.
[online]
Available
from
<http://publications.cancerresearchuk.org/epages/crukstore.sf/en_GB/?ObjectPath=/Shops/crukstore/Categories/BrowseB
ySubject/BreastCancer> [18 September 2004]
Cancer
Research
UK
(2005)
Cancer
stats
incidence-UK.
[online]
Available
<http://www.cancerresearchuk.org/aboutcancer/statistics/statsmisc/pdfs/cancerstats_incidence_apr05.pdf> [10
2005]
from
August
Carole, K. and Sian, L. (2000) 'Knowledge management in the NHS: positioning the healthcare librarian at the knowledge
intersection.' Blackwell Science Ltd. Health Libraries Review 17, 136–143
Cassandra, E.S. (2006) 'Breast cancer screening: cultural beliefs and diverse populations.' Health and social work 31, (1) 36-43
Chen, P., Huang, D., Zhao, X. and Li, X. (2008) ' Predicting contact map using radial basis function neural network with
conformational energy function.' International Journal of Bioinformatics Research and Applications 4, (2) 123-136
Chourides, P., Longbottom, D. and Murphy, W. (2003) ‘Excellence in knowledge management: an empirical study to identify
critical factors and performance measures.’ Measuring Business Excellence 7, (2) 29-45
Chris, W., Folkert, G., Yuan, H.C. (2003) 'Trends in social class differences in mortality by cause 1986 to 2000.’ A National
Statistics Publication UK, Health Statistics Quarterly 20, 25-37
Cicmil, S. (2005) Management of knowledge in project environments. ed: Love, P., Fong, P.S.W. and Irani, Z. Oxford: Elsevier ltd.
155-79
Clarke, K., Rooksby, J., Rouncefield, M., Procter, R., and Slack, R. (2006) 'Healthcare information giving services: technologies
and everyday practicalities.' Health Informatics Journal 12, (2) 153-64
Codd,
A.
(2006)
Withdrawal
of
breast
screening
prior
notification
<http://www.berkshire.nhs.uk/tvpca/documents/TVns/Withdrawal%20of%
20Breast%20Screening%20Prior%20Notification%20Lists.pdf> [10 March 2007]
lists
[online]
Available
from
Collis,J. and Hussey, R. (2003) Business research: A Practical guide for undergraduate and postgraduate students. New York: Palgrave
Macmillan
Davenport, T.H. (1995) ‘Think tank.’ CIO 9, (6) 30-32
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
103
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Day, N.E., Williams, D.D.R., Khaw, K.T. (1989) 'Breast cancer screening programmes: the development of a monitoring and
evaluation system' Br J Cancer 59, 954-8
Day, N.E., McCann, J., Camilleri-Ferrante, C., Britton, P., Hurst, G., Cush, S., Duffy, S. (1995) 'Monitoring intervalcancers in
breast screening programmes: the East Anglian experience.' Journal of Medical Screening 2, 180-185.
Day, K.J., and Norris, A.C. (2007) 'Supporting information technology across health boards in New Zealand: the role of learning
in adapting to complex change.' Health Informatics Journal 13, (3)195-206
DeFillippi, R.J. (2001) ‘Introduction: project-based learning, reflective practices and learning outcomes.’ Management Learning 32
(1) 5-10
Denzin, N.K., (1970) The research act in sociology: a theoretical introduction to scoiological methods. London: Butterworth group
Department of Health (2006) Breast screening- a pocket guide. nhs cancer screening programmes [online] Available from
<http://www.cancerscreening.nhs.uk/ breastscreen/publications/nhsbsp-pocket-guide-2006.pdf> [15 April 2007]
Department of radiology, Brigham and women’s hospital (2007) Breast imaging and diagnostic service. [online] Available from
<http://brighamrad. harvard.edu> [14 July 2007]
Disterer, G. (2002) ‘Management of project knowledge and experiences.’ Journal of Knowledge Management 6, (5) 512-20
Dolin, R.H., Alschuler, L., Boyer, S., Beebe, C., Behlen, F.M., Biron, P.V. and Shabo, A. (2006) ‘HL7 clinical document
architecture, release 2’, Journal of American Medical Informatics Association 13, (1) 30-39
Drucker, P.F. (1993) Post-capitalist society. London: Butterworth Heinemann
Duffy, S.W., Tabar, L., Vitak, B., Yen, M.F., Warwick, J., Smith R.A. and Chen, H.H. (2003) 'The swedish two-county trial of
mammographic screening: cluster randomisation and end point evaluation.' Annals of Oncology [online] 14, 1196–1198. Available
from <http://annonc.oxfordjournals.org/cgi/reprint /14/8/1196.pdf>
Duke, T. (2005) EDI best practice- management guide., Book Industry Communication [online] Available from
<http://www.bic.org.uk/pdf/edi-bestpractice. pdf> [10 August 2005]
Durkheim, E., (1958) The rules of sociological method. ed. by Catlin, G.E.G. University of Chicago: The Free Press
Dwivedi, A., Bali, R.K., James, A.E., Naguib, R.N.G., and Johnston, D (2002) 'Merger of knowledge management and
information technology in healthcare: opportunities and challenges.' Proc of the IEEE Canadian Conference on Electrical and
Computer Engineering (CCECE) 2, 1194-99.
Dwivedi, A., Bali, R.K., Naguib, R.N.G., (2003) 'Organization current knowledge design (ockd): a knowledge management
framework for healthcare institutions.' Proc of the IEEE Ann Int Conf of the Eng in Med and Biol Soc (EMBS) 1236-1239
Eardley, A., and Czerwinski, A. (2007) 'Knowledge management for primary healthcare services.' in healthcare knowledge
management. ed. by Bali RK and Dwivedi A Newyork : Springer, 201-220
Edvinsson, L. (1997) ‘Developing intellectual capital at skandia.’ Long Range Planning 30 (3) 366-73
Eggebraaten, T.J., Tenner, J.W. and Dubbels, J.C. (2007) ‘A health-care data model based on the hl7 reference information
model’, IBM Systems Journal 46, (1) 5-18
Ellingsen, G. (2003) 'The role of trust in knowledge management: a case study of physicians at work at the university hospital
of northern norway.' Informing science journal [online] 6, 193-207. Available from <http ://inform.nu/Articles/Vol6/v6p 193 207.pdf> [24 April 2004]
Epstein, S.S. (1979) The politics of cancer new york, U.S.A: Doubleday, 537
Fernie, S., Green, S.D., Weller, S.J. and Newcombe, R. (2003) ‘Knowledge sharing: context, confusion and controversy.’
International Journal of Project Management 21, 177-87
Fong, P.S.W., ed: Love, P., Fong, P.S.W. and Irani, Z. (2005) Management of knowledge in project environments. Oxford: Elsevier ltd.
103-31
Forrest, P. (1986) Breast cancer screening- A report to the health ministers of England, Scotland, Wales and Northern Ireland. London:
HMSO
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
104
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Fox, S., Murata, P. and Stein, J. (1991) 'The impact of physician compliance on screening mammography for older women.'
Arch. Internal Medicine 151, 50-56
Fraser, H., Pugh, R., Kennedy, R., Ross, P., and Harrison, R. (1994) 'A comparison of backpropagation and radial basis
functions, in the diagnosis of myocardial infarction.' ed. by Ifeachor, E., and Rosen, K. International Conference on Neural
Networks and Expert Systems in Medicine and Healthcare, 76--84
Gillingham, H., and Roberts, B. (2006) 'Implementing knowledge management: a practical approach.' Journal of Knowledge
Management Practice 7 (1) [online] Available from <http://www.tlainc.com/articl107. htm> [10 March 2007]
Glaser, B.G., and Strauss, A.L. (1967) The discovery of grounded theory: strategies for qualitative research. Chicago: Aldine
Goossen, W.T.F., Jonker, M.J., Heitmann, K.U., Hass, I.C.J., Jong, T., Slikke, W. and Kabbes, B.L. (2003) 'Electronic patient
records: domain message information model perinatology.' International Journal of Medical Informatics 70, 265-76
Gray,
M.
(2003)
Knowledge
is
the
enemy
of
disease.
[online]
Available
<http://muirgray.blogspot.com/2003_05_01_muirgray_archive.html accessed 1st April 2005> [23 June 2004]
from
Griswold, B. and Lawrence, S. (2004) Evolving the JRE...from a managed runtime to a manageable runtime. SYS-CON Media [online]
Available from <http://jdj.sys-con.com/read/44372.htm> [15 June 2005]
Guy’s and St Thomas’ Hospital NHS Trust (2006) The new patient transport scheme. [online] Available from
<http://www.guysandstthomas.nhs.uk/ resources/Patient_ Transport_Large_1.pdf> [14 March 2006]
Hackshaw, A. K., Wald N. J., Michell M. J., Field S. and Wilson A.R.M. (2000) 'An investigation into why two-view
mammography is better than one-view in breast cancer screening' Clinical Radiology 55, (6) 454-8
Hall, J. and Sapsed, J., ed: Love, P., Fong, P.S.W. and Irani, Z. (2005) Management of knowledge in project environments. Oxford:
Elsevier ltd. 57-79
Hamilton, E.L., Wallis, M.G., Barlow, J., Cullen, L., Wright, C. (2003) 'Women's views of a breast screening service.' Health Care
for Women International 24, (1) 40- 48
Haykin, S. (1999) Neural networks- a comprehensive foundation. Pearson education
Hersh, W.R. (2002) ‘Medical informatics- improving health care through information.’ Journal of American Medical Association
288, (16) 1955-58
Hersh, W., and Williamson, J. (2007) 'Educating 10,000 informaticians by 2010: the amia 10×10 program.' International Journal of
Medical Informatics 76, 377–82
Hinton, P.R., Broenlow, C., McMurray, I. and Cozens, B. (2004) SPSS explained. UK: Routledge Taylor & Francis Group
publishers
Huang, E., Hsiao, S.H. and Liou, D.M. (2003) ‘Design and implementation of a web-based HL7 validation system.’ International
Journal of Medical Informatics 70, (1) 49-58
Hudson, D.L., Cohen, M.E. (2000) Neural networks and artificial intelligence for biomedical engineering. Series in biomedical
engineering: IEEE press
Huff, S.M. (1998) 'Clinical data exchange standards and vocabularies for messages.' Journal of American Medical Informatics
Association, AMIA Annual Fall Symposium Supplement 62-67
Hussey, T.L., and Hussey, R. (1997) 'Business research: a practical guide for undergraduate and postgraduate students.'
London: Macmillan Press Ltd.
Ichijo, K. and Nonaka, I. (2007) Knowledge creation and management: New challenges for managers. New York: Oxford University
Press
ISO/TC 154 (2002) Electronic data interchange for administration, commerce and transport (EDIFACT) — Application level syntax rules.
UN/CEFACT Syntax Working Group (JSWG) publication of ISO 9735-1 equivalent to the official ISO publication: ISO 9735-1
[online] Available from <http://webstore.ansi.org/ansidocstore/product. asp?sku=ISO+9735-6%3A2002> [10 March 2004]
Jepsen, T. (2003) 'IT in healthcare: Progress report.' IT Professional, IEEE computer society 5, (1) 8-14.
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
105
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Jepson, R., Clegg, A., Forbes, C., Lewis, R., Sowden, A. and Kleijnen, J. (2000) ‘The determinants of screening uptake and
interventions for increasing uptake: a systematic review.’ Health Technology Assess 4, (14)
Joint ISO/TC 154 – UN/CEFACT Syntax Working Group (JSWG) (2002) Electronic data interchange for administration, commerce
and transport (EDIFACT) — Application level syntax rules. [online] publication of ISO 9735-1 equivalent to the official ISO
publication: ISO 9735-1 Available from <http://webstore.ansi.org/ansidocstore/product.asp?sku=ISO+9735-6%3A2002> [20
April 2005]
Joos, I.M. (2005) Introduction to computers for healthcare professionals. US: Jones & Bartlett
Kanungo, T., Mount, D.M., Netanyahu, N.S., Piatko, C.D., Silverman, R., and Wu, A.Y. (2002) ' An efficient k-means clustering
algorithm: analysis and implementation.' Pattern Analysis and Machine Intelligence, IEEE Transactions 24, (7) 881-892
Kasvi, J.J.J., Vartiainen, M., and Hailikari, M. (2003) ‘Managing knowledge and knowledge competences in projects and project
organisations.’ International Journal of Project Management 21, 571-82
Katz, S.J., Zemencuk, J.K., Hofer, T.P. (2000) 'Breast cancer screening in the United States and Canada, 1994: Socioeconomic
Gradients Persist.' American Journal of Public Health 90, (5) 799-803
Kelly, (1999), 'System directed interventions', UK, Berkshire, un-published work reported by Sin, J.P. and Leger, A.S. in
'Interventions to increase breast screening uptake: do they make any difference?' Journal of Medical Screening 6, (1) 170-181
Koskinen, K.U. (2003) 'Evaluation of tacit knowledge utilization in work units.' Journal of Knowledge management 7, (5) 67-81
Kurzweil, R. (1992) The age of intelligent machines: can machines think? The MIT Press
Kwon, D. (2004) Knowledge management for turbulent times: performance impacts of IT valuation. Quebec City, Canada: ASAC
Conference
Lagerlund, M., Hedin, A., Spare´n, P., Thurfjell, E., Lambe, M. (2000) 'Attitudes, beliefs, and knowledge as predictors of
nonattendance in a swedish population-based mammography screening program.' Journal of Preventive Medicine 31, 417–428
Lagerlund, M., Maxwell, A.E., Bastani, R., Thurfjell, E., Ekbom, A., Lambe, M. (2002) 'Sociodemographic predictors of nonattendance at invitational mammography screening – a population-based register study (Sweden).' Journal of Cancer Causes and
Control 13, 73–80
Lavra, N. et al. (2003) Machine learning. ECML 2003: 14th European Conference on Machine Learning. Cavtat-Dubrovnik:
Springer
Learning To Learn (2006) Structure of a typical neuron. [online] Available from <http://www.learningtolearn.sa.edu> [10
August 2005]
Lehaney, B., Jack, G., Clarke, S., and Coakes, E. (2003) Beyond knowledge management. Hershey: Idea Press
Lelic, S. (2002) 'Your Say: KM in the healthcare industry.' Knowledge Management Magazine 6, (4)
Liebowitz, J. (1999) The knowledge management handbook. Boca Raton, FL: CRC press
Liebowitz, J. (2005) Management of knowledge in project environments. ed. by Love, P., Fong, P.S.W. and Irani, Z. Oxford: Elsevier
ltd. 1-18
Love, P.E.D., Huang, J., Edwards, D.J. and Irani, Z., ed: Love, P., Fong, P.S.W. and Irani, Z. (2005) Management of knowledge in
project environments. Oxford: Elsevier ltd. 133-54
Lytras, M.D. and Pouloudi, A. (2003) ‘Project management as a knowledge management primer: The learning infrastructure of
knowledge-intensive organizations: Projects as knowledge transformations and beyond.’ The learning organization 10, (4) 237-50
MacClelland, J.L., Rumelhart, D.E. (1986) Parallel distributed processing: explorations in the microstructure of cognition. MIT Press
MacIntyre, J., Moscardini, A. and Phillips,P. (2002) 'Improving the classification ability of neural networks using exponential
smearing on clustered data' Artificial and Computational Intelligence 365
Maglogiannis, I., Sarimveis, H., Kiranoudis, C.T., Chatziioannou, A.A., Oikonomou, N., Aidinis, V. (2008) 'Radial basis function
neural networks classification for the recognition of idiopathic pulmonary fibrosis in microscopic images.' IEEE Trans Inf
Technol Biomed 12, (1) 42-54
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
106
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Maguire, S., and Ojiako, U. (2007) 'Interventions for information systems introduction in the NHS.' Health Informatics Journal
13,(4) 283–302
Majeed, A., Given-Wilson, R. and Smith, E. (1997) 'Impact of follow up letters on non-attenders for breast screening: a general
practice based study.' Journal of Medical Screening 4, 19-20
Majeed, F.A., Cook, D.G., Given-Wilson, R., Vecchi, P., Poloniecki, J. (1995) 'Do general practitioners influence the uptake of
breast cancer screening?' Journal of Medical Screening 2, 119-24.
Mandl, K.D., Szolovitis, P. and Kohane, I.S. (2001) ‘Public standards and patients’ control: how to keep electronic medical
records accessible but private’, BMJ 322, :283-7
Marcela, B.V. (2004) 'The system does work.' Journal of American College of Radiology 1, (6) 438-40
Margarita, S. (2002) Introduction to neural networks in healthcare. Open Clinical- KM for medical care. [online] Available from
<http://www. openclinical.org/docs/int/neuralnetworks011.pdf> [10 March 2005]
Markwel, D. (1999) Strategic review of healthcare message standards alignment - draft final report v0.5, CEN/TC 251/N99-093,
CEN/TC
251
Secretariat:
Swedish
Healthcare
Standards
Institution
[online]
Available
from
<http://www.centc251.org/TCMeet/Doclist/TCdoc99 /N99-093.rtf> [10 March 2004]
Marr, B. (2004) ‘Measuring and benchmarking intellectual capital.’ Benchmarking: an International Journal 11, (6) 559-70
Martin, C. (2005) UK NHS diabetic retinopathy message specifications – guidance notes Release 3.0, 2005-09-28.' [online] Available
from
<http://www.nscretinopathy.org.uk/resources/Message%20
specification%
20guidance%20notes%20%20Release%203.0%202005-09-28.doc> [13 March 2006]
Martin. D., Mariani. J., and Rouncefield. M. (2007) 'Managing integration work in an NHS electronic patient record (EPR)
project.' Health Informatics Journal 13, (1) 47-56
Maslow, A. (1943) ‘A theory of human motivation.’ Psychological Review 50, 370-96
Masters, T. (1995) Practical neural network recipes in C++. New York: John Wiley Sons
Mc Donald. C.J., (1997) 'The barriers to electronic medical record systems and how to overcome them.' Journal of the American
Medical Informatics Association 4, (3) 213-21
McCarthy, J., Minsky, M.L., Rochester, N. and Shannon, C.E. (1955) A proposal for the dartmouth summer research project on
artificial intelligence. [online] Available from <http://www-formal.stanford.edu/jmc/history/ dartmouth/dartmouth.html> [20
April 2005]
McDermott, R. (1999) ‘Why information technology inspired but cannot deliver knowledge management.’ Claifornia
Management Review 41, (4) 103-17
McElroy, M.W. (2003) The new knowledge management. Butterworth-Heinemann publications: KMCI press, 13-18
McRoy, S.W., Liu-Perez, A. and Ali, S.S. (1998) 'Interactive computerized health care education.' Journal of the American Medical
Informatics Association 5, 347-356
Medicine net (2002) Breast cancer. [online] Available from <http://www.medicinenet.com/breast_cancer/page3.htm> [10
August 2005]
Melville, S.K., Luckmann, R., Coghlin, J., Gann, P. (1993) 'Office systems for promoting screening mammography: a survey of
primary care practices.' Journal of Family Practice 37, (6) 569-74
Miller, F.J. (2002) 'I = 0 (Information has no intrinsic meaning).' Information Research - International electronic journal [online] 8,
(1). Available from <http ://informationr.net/ir/8-1/paper 140.html> [08 October 2002]
Milner, B. (2001) 'Knowledge management: the challenge of the twenty-first century.' Problems of economic transition 43, (10) 6578
Milton, N. (2008) Assigning roles and accountabilities in km. knowledge management review [online] Available from
<http://findarticles.com/p/ articles/mi_qa5362/is_200803/ai_n25418486>
Moor, J.H. (2003) The turing test: the elusive standard of artificial intelligence. Springer
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
107
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
National
Cancer
Institute
(2005)
The
breasts.
<http://www.cancer.gov/cancertopics/wyntk/breast/page2> [10 August 2005]
[online]
Available
from
Negnevitsky, M. (2004) Artificial intelligence: a guide to intelligent systems. US: Addison Wesley
Newell, S. and Huang, J., ed: Love, P., Fong, P.S.W. and Irani, Z. (2005) Management of knowledge in project environments. Oxford:
Elsevier ltd. 19-39
Newman, I., and Benz, C.R. (1998) 'Qualitative-quantitative research methodology: Exploring the interactive continuum.'
Carbondale, IL: Southern Illinois University press
NHS (2004) NeLH [online] Available from <http ://www.nelh.nhs.uk/knowledge_management/kml /nhs.asp>
NHS
CfH
resources
(2007)
NHS
Connecting
for
Health
fact
Sheet.
<http://www.connectingforhealth.nhs.uk/resources/sys tserv /nhscfh> [16 may 2007]
[online]
NHS
Connecting
For
Health
(2007a)
The
N3
fact
sheet.
[online]
<http://www.connectingforhealth.nhs.uk/systemsandservices/n3/factsheet> [28 January 2008]
Available
Available
from
from
NHS Connecting For Health (2007b) The SPINE fact sheet. [online] Available from <http://www.connectingforhealth
.nhs.uk/resources/systserv/spine-factsheet> [28 January 2008]
NHS
Connecting
for
Health
Fact
Sheet
(2007)
NHS
CfH
resources.
<http://www.connectingforhealth.nhs.uk/resources/ systserv/nhscfh> [16 May 2007]
[online]
Available
from
NHS Information Standards Board (2006) Information standards board for health and social care. [online] Available from
<http://www.isb.nhs.uk> [16 may 2007]
NHS
IT
Standards
Handbook
(2005)
Part
2Chapter
225<http://www.hl7italia.it/bac_Utenticoordinatori_717/ 225chap.pdf> [16 may 2007]
HL7.
[online]
Available
NHS
Review
(2006)
The
breast
screening
programme
annual
review
2005.
[online]
<http://www.cancerscreening.nhs.uk/breastscreen// publications/ 2005review.html> [14 July 2007]
NHS
Review
(2007)
The
breast
screening
programme
annual
review
2006.
[online]
<http://www.cancerscreening.nhs.uk/breastscreen/ publications /2006review.html> [14 July 2007]
NHSBSP
(1999-2005)
NHS
breast
screening
programme
annual
<http://cancerscreening.org.uk/breastscreen/publications/> [14 March 2006]
review.
[online]
from
Available
Available
Available
at
from
from
Nonaka, D., Teece, D.J. Kikawada, K. and Holtshouse D (2001) Managing industrial knowledge. Sage publications ltd. 306-314
Nonaka, I. and Takeuchi, H. (1995) The knowledge-creating Company. USA: Oxford University Press 61-94
Noseworthy. T. [2004] 'Advancing electronic health records in canada: why, how and key learnings of potential value to china.'
IEEE Computer Society, Medical Information Systems: The Digital Hospital, IDEAS Workshop 129-32
O’Connor, A., Griffiths, C., Underwood, M. and Eldridge, S. (1998) ‘Can postal prompts from general practitioners improve the
uptake of breast screening? A randomised controlled trial in one east London general practice.’ Journal of Medical Screening 5,
49-52
OASIS (2004) Case study: UK National Health Service NPfIT uses ebXML messaging
<http://www.ebxml.org/case_studies/ NHS-ebMSG-casestudy-041206.pdf > [12 May 2005]
[online]
Available
from
Oikonomou, A., Amin, S.A., Naguib, R.N.G., Todman, A., Al-Omishy, H. (2003) 'Breast self examination training through the
use of multimedia: a prototype multimedia application' Engineering in Medicine and Biology Society, Proceedings of the 25th
Annual International Conference of the IEEE 2, (21) 295-98
Òkane,
K.
(2002)
Mumps
language
bioinformatic
database
resources.
<http://bioinformatics.org/forums/forum.php?forum_id =1035> [20 April 2005]
[online]
Available
from
Oliver. K.B., and Roderer. N.K. (2006) 'Working towards the informationist.' Health Informatics Journal 12,(1) 41-8
Open Clinical (2000) The medical knowledge crisis and its solution through knowledge management- White Paper - (DRAFT- v.3-15).
[online] Available from <http: //www. openclini cal.org/docs/whitepaper.pdf> [10 March 2004]
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
108
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Osborna, G.D., Beera, H., Wadeb, R., Brooka, D., Stevensa, G., Evansa, J., Fielder, H. and Gower-Thomas, K.(2006) 'Two-view
mammography at the incident round has improved the rate of screen-detected breast cancer in Wales.' Clinical Radiology 61, (6)
478-482
Pam,
S.
(2007)
Mammogram
views
for
routine
and
diagnostic
screening.
<http://breastcancer.about.com/od/mammograms/ a/mamm_views.htm> [14 July 2007]
[online]
Available
from
Patnick, J. (2004) 'NHS breast screening: the progression from one to two views, Journal of Medical Screening 11, (2) 55-56
Patnick J. et al. (2006) 'Screening for breast cancer in England: past and future.' Journal of medical screening 13, 59-61
Patnick, J. (2006) Disclosure of audit results in cancer screening advice on best practice, Cancer screening series 3. NHS Cancer
Screening Programmes [online] Available from <http://www.cancerscreening.nhs.uk/publications/cs3.pdf> [16 May 2007]
Paul, H. and John, P. (1997) 'Networking and data communications basics.' CPA journal [online] Available from
<http://www.nysscpa.org/cpajournal /1997/1197/features/f461197.htm> [10 August 2005]
Paul, M.H. and Kimble, C. (2002) 'The duality of knowledge.' Information Research- International electronic journal. [online] 8, (1).
Available from <http ://informationr.net/ir/8-1/paper 142.html> [10 March 2004]
PCI (2007) Postcode distance calculator [online] available from <http://www.pc-i.co.uk/postcode-distance.php> [16 may 2007]
Pemberton, J.D., Stonehouse, G.H. and Yarrow, D.J. (2001) ‘Benchmarking and the role of organizational learning in developing
competitive advantage.’ Knowledge and Process Management 8, (2) 123-35
Pemberton, J.M. (1998) 'Knowledge management (KM) and the epistemic tradition.' ARMA Records Management Quarterly 32, (3)
58-62
Pirjo, I., Kauhava, L., Parvinen, I., Helenius, H., Klemi, P. (2001) 'Customer fee and participation in breast cancer screening.' The
Lancet 358, 1425
Pisano, E.D., Cole, E.B., Kistner, E.O., Muller, K.E., Hemminger, B.M., Brown, M.L., Johnston, R.E., Kuzmiak, C.M., Braeuning,
M.P., Freimanis, R.I., Soo, M.S., Baker, J.A., Walsh, R. (2002) 'Interpretation of digital mammograms: comparison of speed and
accuracy of soft-copy versus printed-film display.' Radiology 223:482-488
Pizziferri, L., Kittler, A.F., Volk, L.A., Honour, M.M., Gupta, S., Wang, S., Wang, T., Lippincott, M., Li, Q. and Bates, W. (2005)
‘Primary care physician time utilization before and after implementation of an electronic health record: A time-motion study.’
Journal of Biomedical Informatics 38, 176-88
Plucker,
J.
(2007)
Human
intellgence-historical
foundations.
<http://www.indiana.edu/~intell/periodIndex.shtml> [23 December 2007]
[online]
Available
from
Polanyi, M. (1961) Knowing and being, Mind. New Series 70, (280) 458-70
Polanyi, M. (1966) The tacit dimension. Garden city, New York: Doubleday and co.
Prasad, B. (2001) ‘Total value management- a knowledge management concept for integrating TQM into concurrent product
and process development.’ Knowledge and Process Management 8, (2) 105-22
Prather, J.C., Lobach, D.F., Goodwin, L.K., Hales, J.W., Hage, M.L. and Hammond, W.E. (1997) 'Medical data mining:
knowledge discovery in a clinical data warehouse.' Proc AMIA Annu Fall Symp. 101-5
Prencipe, A., Brady, T., Marshall, N. and Tell, F., ed: Love, P., Fong, P.S.W. and Irani, Z., (2005) Management of knowledge in
project environments. Oxford: Elsevier ltd. 197-217
Project Management Institute (2004) A guide to the project management body of knowledge (PMBOK® Guide) 3rd edition. USA: PMI
Protti, D.J. (2005) The application of computer technology in GP surgeries is beginning to have positive effects on chronic disease
management. Prof. DJ Protti’s World View Reports, NHS: CfH Press office [online] Available from
<http://www.connectingforhealth.nhs.uk /newsroom/world view/protti> [20 April 2005]
Prusak, L. (2001) 'Where did knowledge management come from?’ IBM Systems Journal 40, (4) 1002-7
Prusak, L. and Matson, E. (2006) Knowledge management and organizational learning. Oxford: Oxford University Press
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
109
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Report by the comptroller and auditor general, department of health (2006) The National Programme for IT in the NHS. healthcare
1173, Session 2005-2006. [online] Available at <http://www.nao.org.uk/publications/nao_reports/05-06/05061173.pdf> [28
January 2008]
Richards, S.H., Bankhead, C., Peters, T.J., Sharp, D.J., Hobbs, F.D.R., Brown, J., Roberts, L., Tydeman, C., Redman, V., Formby,
J., Wilson, S. and Austoker, J. (2001) 'Improving attendance for breast screening among recent non-attenders: a randomised
controlled trial of two interventions in primary care.' Journal Medical Screening 8, 99-105
Rizvan, E., Seyfettin, N.O., Cenk, Ş. and Nazan, Z.A. (2008) 'A radial basis function neural network (RBFNN) approach for
structural classification of thyroid.' Journal of Medical Systems 32, (3) 215-220
Rodrigo, B.C. (2001) 'Using information technology to support knowledge conversion processes.' Information Research International electronic journal [online] 7:1. Available from <http://informationr.net/ir/7-1/paper118.html> [10 August 2005]
Royal College of GP (RCGP) (2005) General practice in the uk:
<http://www.rcgp.org.uk/pdf/ISS_INFO _04_MAY05.pdf> [23 June 2006]
a
basic
overview.
[online]
Available
from
Russell, A.S. (2000) Demystifying EDI. Plano, TX : Wordware Pub.
Ryan, A. S. (2006) ‘Towards semantic interoperability in healthcare: Ontology mapping from SNOMED-CT to HL7 version 3’,
In Proc. Second Australasian Ontology Workshop (AOW 2006), Australia. CRPIT, ACS 72, 69-74
Saywell, R.M., Champion, V.L., Skinner, C.S., McQuillen, D., Martin, D., Maraj, M. (1999) 'Cost–effectiveness comparison of
five interventions to increase mammography screening.' Preventive Medicine 29, 374–382
Schlegel,
S.
(2003)
ebXML,
an
open
standard.
<http://www.schlegel.li/ebXML/linmagau_article/linmagau.pdf> [10 March 2004]
[online]
Available
from
Schwenk, C.R. (1988) ‘The cognitive perspective on strategic decision making.’ Journal of management studies 25, (1) 41-55
Senge, P. (1990) The fifth discipline: the art and practice of the learning organisations. New York: Century Business
Sharkie, R. (2003) 'Knowledge creation and its place in the development of sustainable competitive advantage.' Journal of
Knowledge management, MCB University press, 7, (1) 20-31
Shea, S., DuMuchel, W. and Bahamonde, L. (1996) 'A meta analysis of 16 randomized controlled trials to evaluate computerbased clinical reminder systems for preventive care in the ambulatory setting.' Journal of American Medical Informatics Association
3, (6) 399-409
Shindler, M. and Eppler, M.J. (2003) ‘Harvesting project knowledge: A review of project learning methods and success factors.’
International Journal of Project Management 21, 219-28
Shinji, S., Shingo, I., Tok, W.L., Kwan-Ho S (Editors) (2005) Web and communication technologies and internet-related social issues HSI 2005. Springer
Siew, E.G., Smith, K., Churilov,L. and Ibrahim, M. (2002) 'A neural clustering approach to iso-resource for acute healthcare in
australia.' Proceedings of the 35th Annual Hawaii International Conference on System Sciences (HICSS'02) 6, (6) 154
Sin, J.P. and Leger, A.S. (1999) 'Interventions to increase breast screening uptake: do they make any difference?' Journal of
Medical Screening 6, (1) 170-181
Sinnott, R. (1984) “Virtues of the Haversine.” Sky & Telescope, 68 (2), 159
Skaane, P. and Skjennald, A. (2004) 'Screen-film mammography versus full-field digital mammography with soft-copy reading:
randomized trial in a population-based screening program--the Oslo II Study.' Radiology 232, (1) 197-204
Skaane, P., Young, K., Skjennald, A. (2003) 'Population-based mammography screening: comparison of screen-film and fullfiled digital mammography with soft-copy reading-oslo 1 study.' Radiology 229, 877-884
Smith, L. and Preston, H. (2000) ‘Information management and technology strategy in healthcare: local timescales and national
requirements’, Information Research- International electronic journal [online] 5, (3) Available from <http://informationr.net/ir/5 3/paper74 .html> [10 March 2004]
Smith. B.,and Ceustersc. W. (2006) 'HL7 RIM: an incoherent standard.' Studies in Health Technology and Informatics 124, 133–8
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
110
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Smith-Bindman, R.,Ballard-Barbash, R., Miglioretti, D.L., Patnick, J., Kerlikowske, K. (2005) 'Comparing the performance of
mammography screening in the USA and the UK.' Journal of Medical Screening 12, (1) 50-54
Spahni, S., Lovis, C., Mercille, R., Verdel, H., Cotton, M. and Geiddbühler, A. (2007) 'Implementing a new ADT based ion the
HL7 version 3 RIM.' International Journal of Medical Informatics 76, (2-3) 190-4
Spyrou, S.S., Bamidis, P., Chouvards, I., Gogou, G., Tryfon, S.M.and Maglaveras, N. (2002) ‘Healthcare information standards:
comparison of the approaches’, Health Informatics Journal 8, 14-9
Stead, M.J., Wallis, M.G., Wheaton, M.E. (1998) 'Improving uptake in non-attenders of breast screening: selective use of second
appointment.' Journal of Medical Screening 5, (2) 69-72
Stonehouse, G.H. and Pemberton, J.D. (1999) ‘Learning and knowledge management in the intelligent organisation,
participation and empowerment.’ An International Journal 7, (5) 131-44
Suliman, A.H. (2002) ‘Knowledge management: re-thinking information management and facing the challenge of managing
tacit knowledge’, Information Research - International electronic journal [online] 8, (1) Available from
<http://informationr.net/ir/8-1/paper 143 .html> [10 March 2004]
Sveiby, K. (1998) Intellectual capital: Thinking ahead. Australian CPA 68, (5) 18-21
Swan, J., Newell, S., Scarborough, H. and Hiscop, D. (1999) ‘Knowledge management and innovation: networks and
networking.’ Journal of knowledge management 3, (4) 262-75
Szczepura A (2005) South Asian population less likely to attend cancer screening than non-asian. [online] Available from
<http://www2.warwick.ac.uk/ newsandevents/pressreleases/south_asian_population/>
Tabar, L., Gad, A., Holmberg, L.H. et al. (1985) 'Reduction in mortality from breast cancer after mass screening with
mammography.' Lancet 1, 829-32
Tabar, L., Vitak, B., Chen, H.H. et al. (2000) 'The Swedish two-county trial twenty years later. Updated mortality results and
new insights from long term follow-up.' Radiologic clinics of North America 38, 625-51
Tarver, S., Cronin-Cowan, K. and Wheaton, M.E. (2004) 'A pilot's life for us, Symposium Mammographicum 2004.' Breast Cancer
Res 6, (1) 52
Taylor, R., Ivanov, O., Brotherton, J. and Achat, H. (2003) 'Predictors of non-attendance from Breast Screen NSW in women
who report current mammography screening.' Australian and New Zealand journal of public health 27, (6) 581-7
The European Commission within the Sixth Framework Programme (2006) RIDE- a roadmap for interoperability of ehealth systems
in support of COM 356 with special emphasis on semantic interoperability. [online] Available from <http://www.srdc.metu.edu.tr
/webpage/projects/ride/deliverables/ RIDE-_3_2_1Vision-v1.4INTERIM.doc> [16 may 2007]
The Information Centre (2007) Trends in consultation rates in general practice 1995 - 2006. [online] Available from
<http://www.ic.nhs.uk/statistics-and-data-collections/primary-care/general-practice/trends-in-consultation-rates-ingeneral-practice-1995--2006> [02 February 2008]
Thomas, J.K., Kellogg, W.A. and Erickson, T. (2001) ‘The knowledge management puzzle: human and social factors in
knowledge management.’ IBM Systems Journal 40, (4) 863-84
Thorp, J. (2005) Secondary uses service strategic direction. SUS Project Board, Connecting for Health [online] Available from
<http://www.connectingforhealth.nhs.uk/systemsandservices/sus/ reference/sus_vision.pdf> [20 June 2006]
Timo, K. (2001) Knowledge management process model. Technical Research Centre of Finland: VTT Publications 455, 101-04
Tissen, R., Andriessen, D. and Deprez, F.L. (2000) The knowledge dividend. Financial Times: Perntice hall, 184-202
Tsoukas, H. and Vladmiron, E. (2001) ‘What is organizational knowledge.’ Journal of management Studies- Special issue: KM
concepts and controversies 38, (7) 973-94
Turner, K., Wilson, J. and Gilbert, J. (1994) ‘Improving breast screening uptake: persuading initial non-attenders to attend.’
Journal of Medical Screening 1, 199-202
UK EDIFACT (1998) Trade message convention. Article Number Association- ana5012345678900 [online] Available from
<http://www.gefeg.com/public /practice/ana/index.htm>
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
111
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
United Nations Economic Commission for Europe, (2006) A roadmap towards paperless trade. New York and Geneva: UN
Vainio, H., Bianchini, F. (eds) (2002) Breast cancer screening, International Agency of Research on Cancer (IARC) Handbooks of Cancer
Prevention. Lyon: IARC Press
Verma, V.K. (1995) Organization projects for success. USA: Project Management Institute, vol1
Vishwanath. A., and Scamurra. S.D. (2007) 'Barriers to the adoption of electronic health records: using concept mapping to
develop a comprehensive empirical model.' Health Informatics Journal 13, (2) 119-134
Von Lubitz , D. and Wickramasinghe, N. (2006) ‘Creating germane knowledge in dynamic environments.’ International Journal
of Innovation and Learning (IJIL) 3, (3) 326-347
Walsham, G. (2001) ‘Knowledge management: The benefits and limitations of computer systems.’ European Management Journal
19, (6) 599-608
Walter, E. (2005) Cambridge advanced learner’s dictionary. Cambridge : Cambridge University Press, 2nd edn.
Ward, R. (2003) The e-GIF and the NHS - a policy statement. Department of Health [online] Available from
<http://www.dh.gov.uk/en/Publicationsand statistics/Publications/PublicationsPolicyAndGuidance/DH_4130864> [10
March 2005]
Wasserman, P.D. (1993) Advanced methods in neural computing. New YorkUSA: John Wiley & Sons, Inc.
Wernerfelt, B. (1984) ‘A resource-based view of the firm.’ Strategic Management Journal 5, (2)171-80
Whelton, M., Ballard, G. and Tommelein, I.D. (2002) 'A knowledge management framework for project definition.' IT con 7, 197212
Wickramasinghe. N., Bali. R.K. and Geisler. E. (2007) The major barriers and facilitators for the adoption and implementation of
knowledge management in healthcare operations, International Journal of Electronic Healthcare 3, (3) 367-381
Williams,
C.
(2005)
Standards
development
guide.
PISCES
<http://www.pisces.co.uk/docs/pisces/wgstandards.pdf> [15 May 2006]
Limited
[online].
Available
from
Wilson, C. (1999) Coverage and uptake in the NHS breast screening program. MSc Project, Coventry: Coventry University
Wilson, T.D. (2002) ‘The nonsense of knowledge management.’ Information Research- International electronic journal [online] 8, (1)
Available from <http://informationr.net/ir/8-1/paper144.html> [10 August 2005]
Wolfgang, K. and Uwe, K. (2001) Influence of the Internet on data collection and dissemination in the European statistical system (ESS).
IAOS
Satellite
Meeting
on
Statistics
for
the
Information
Society
[online]
Available
from
<http://www.stat.go.jp/english/info/meetings/iaos /pdf/knuppel.pdf> [20 April 2005]
Yin, R.k. (1994) Case study research:design and methods. London: Sage publications
Zajtchuk, R., Satava, R.M. (1997) Medical applications of virtual reality. Communications of the ACM archive, ACM Press 40, (9)
63-4
Zorbas, H.M (2003) 'Breast cancer screening.' MJA 178, (12) 651-652
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
112
Implementation of a breast cancer screening prediction algorithm: a knowledge management approach
A PhD project funded by the NHS Cancer Screening Programmes
Carried out by the Biomedical Computing and Engineering Technologies (BIOCORE) Applied Research Group, Coventry
University (www.coventry.ac.uk/biocore/)
Appendix:
Research Papers Published
1.
Baskaran V, Bali RK, Shah B and Guergachi A, (2009) "Knowledge management as a holistic tool for superior project
management”, International Journal of Innovation and Learning by InderScience,ISSN: 1471-8197, (forthcoming)
2.
Baskaran V, Bali RK, Arochena H, Naguib RNG, Wheaton M, Wallis M , Benson T and Wickramasinghe N (2009) "A
knowledge-based primary care approach to increase breast screening attendance", Int. J. of Biomedical Engineering
and Technology, 2(2) 172-88
3.
Baskaran V, Bali RK, Arochena H, Naguib RNG, Wheaton M, Wallis M , Benson T and Wickramasinghe N (2009)
“Physician intervention via Knowledge Management: using HL7 messaging to increase breast screening uptake”
(forthcoming)
4.
Baskaran V, Bali RK, Arochena H, Naguib RNG, Khanapi MAG, Shah B, Guergachi A, Byrne C (2008) “Health
informatics: a UK and Canadian perspective on current and future challenges”, The 14th Americas Conference on
Information Systems, AMCIS 2008, Toronto, Ontario, pp. 204 -004
5.
Khanapi MAG, Bali RK, Naguib RNG, Marshall IM, Baskaran V, Puentes J, Wickramasinghe NS (2008) "A flexible
telemedicine framework for the continuous upkeep of patient lifetime health records (f2u-lhr)", The 14th Americas
Conference on Information Systems, AMCIS 2008, Toronto, Ontario, pp. 204-005
6.
Baskaran V, Bali RK, Arochena H, Naguib RNG, Wallis M and Wheaton M (2006) "Knowledge creation using
artificial intelligence: a twin approach to improve breast screening attendance", Proc of the 2006 IEEE Eng in Med and
Biol 28th Ann Conf, New York City, USA, 30 Aug - 3 Sept 2006, pp. 4070 - 4073
7.
Baskaran V, Bali RK, Arochena H, Naguib RNG, Wheaton M and Wallis M (2006) "Improving uptake of a breast
screening programme: a knowledge management approach for opportunistic intervention" In Bos L et al (Eds)
Medical and Care Compunetics 3, IOS:Netherlands, pp.191-197
8.
Baskaran V, Bali RK, Arochena H, Naguib RNG (2005) “A knowledge management approach to increase uptake in a
breast screening programme”, Proceedings of the Second Humanoid, Nanotechnology, Information Technology,
Communication and Control Environment and Management (HNICEM) International Conference, The Institute of Electrical
and Electronics Engineers Inc. (IEEE) – Philippine Section. 17-20 March 2005, pp. 1173- 1177
9.
Baskaran V, Bali RK, Arochena H, Naguib RNG, Dwivedi AN and Nassar NS (2004) "Towards total knowledge
management for healthcare: clinical and organizational considerations", Proc of the IEEE-EMBC 26th Ann Int Conf of
the IEEE Eng in Med and Biol Soc (EMBS), September 2004, pp. 3163-3166, San Francisco, USA, pp. 3163- 3166
10.
Dwivedi AN, Bali RK, Naguib RNG, Baskaran V and Nassar NS (2003) "Tracking the healthcare management
landscape: implications for integrated healthcare management and practice", Proc of the IEEE-EMBC 25th Ann Int Conf
of the IEEE Eng in Med and Biol Soc (EMBS), 17-21 September 2003, Cancun, Mexico, pp. 1240-1243
All enquiries relating to this report should be sent to Dr Rajeev K Bali, r.bali@ieee.org
113