Full Chapter Development of Clinical Decision Support Systems Using Bayesian Networks With An Example of A Multi Disciplinary Treatment Decision For Laryngeal Cancer Mario A Cypko PDF
Full Chapter Development of Clinical Decision Support Systems Using Bayesian Networks With An Example of A Multi Disciplinary Treatment Decision For Laryngeal Cancer Mario A Cypko PDF
Full Chapter Development of Clinical Decision Support Systems Using Bayesian Networks With An Example of A Multi Disciplinary Treatment Decision For Laryngeal Cancer Mario A Cypko PDF
https://textbookfull.com/product/the-multi-criteria-approach-for-
decision-support-an-introduction-with-practical-applications-
lotfi-azzabi/
https://textbookfull.com/product/decision-support-systems-viii-
sustainable-data-driven-and-evidence-based-decision-support-
fatima-dargam/
https://textbookfull.com/product/deep-learning-for-medical-
decision-support-systems-utku-kose/
https://textbookfull.com/product/vba-for-modelers-developing-
decision-support-systems-with-microsoft-office-excel-5th-edition-
s-christian-albright/
Analytics, data science, & artificial intelligence :
systems for decision support Eleventh Edition Dursun
Delen
https://textbookfull.com/product/analytics-data-science-
artificial-intelligence-systems-for-decision-support-eleventh-
edition-dursun-delen/
https://textbookfull.com/product/deep-learning-in-medical-image-
analysis-and-multimodal-learning-for-clinical-decision-support-
danail-stoyanov/
https://textbookfull.com/product/real-world-decision-support-
systems-case-studies-1st-edition-jason-papathanasiou/
https://textbookfull.com/product/optimization-and-decision-
support-systems-for-supply-chains-1st-edition-ana-paula-barbosa-
povoa/
Mario A. Cypko
Development of
Clinical Decision
Support Systems
using Bayesian
Networks
With an example of a Multi-
Disciplinary Treatment Decision for
Laryngeal Cancer
Development of Clinical Decision
Support Systems using Bayesian
Networks
Mario A. Cypko
Development of
Clinical Decision
Support Systems
using Bayesian
Networks
With an example of a Multi-
Disciplinary Treatment Decision for
Laryngeal Cancer
Mario A. Cypko
Innovation Center Computer Assisted Surgery,
Medical Faculty
University of Leipzig
Leipzig, Germany
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Fachmedien
Wiesbaden GmbH, part of Springer Nature 2020
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher,
whether the whole or part of the material is concerned, specifically the rights of translation,
reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any
other physical way, and transmission or information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are
exempt from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in
this book are believed to be true and accurate at the date of publication. Neither the publisher
nor the authors or the editors give a warranty, expressed or implied, with respect to the material
contained herein or for any errors or omissions that may have been made. The publisher remains
neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This Springer Vieweg imprint is published by the registered company Springer Fachmedien Wiesbaden
GmbH part of Springer Nature.
The registered company address is: Abraham-Lincoln-Str. 46, 65189 Wiesbaden, Germany
”If I have seen further, it is by standing on the shoulders of giants.”
— Isaac Newton, 1676
Acknowledgments
This thesis combines multiple technical and clinical disciplines and never could have
reached this level without the support and many inspiring discussions with giants of var-
ious professions. Those who are not personally named here but crossed my scientific
journey, you influenced my work as well, thank you.
Personally, first and foremost, I would like to thank my supervisors Prof. Heinz U.
Lemke and Prof. Andreas Dietz as well as my thesis supporter Prof. Gerhard Brewka.
Heinz Lemke was the visionary of ICCAS’s digital patient model project and senior ad-
visor of the Digital Patient and Process Model group. Dear Heinz, thank you for sharing
your knowledge, experiences and visions. Thank you for your belief in a raw diamond and
your effort in grinding it. Thank you for your friendship. Prof. Dietz, thank you and your
clinical team for the openness, unlimited support and considering me as a part of your clin-
ical family. I sincerely thank Prof. Brewka from the Computer Science Department of the
University of Leipzig for supporting me in the writing and submission process of this thesis.
From the clinical team, I would like to emphasize my dear clinical colleague Dr. Matthaeus
Stoehr. Matthaeus, thank you for four years of intensive and ambitious teamwork. None of
our research would ever reach the quality without your clinical and personal influence.
I thank the Innovation Center Computer Assisted Surgery of the Medical Faculty of
the University of Leipzig and specifically my both team leaders during the time of my
research, Dr. Kerstin Denecke and Dr. Steffen Oeltze-Jafra. Your trust in me and my
project strategies enabled a full development of my creativity. Furthermore, I would like to
acknowledge my colleagues Yihan Deng and Jan Gaebel as well as my student assistants
who supported this work and implemented the tools of the presented work, especially
Stefanie Schlinke, David Hirsch, Janine Brach, and Lara Heuft, thank you. I also thank the
Federal Ministry of Research and Education for funding my research and travel expenses
as well as endorsing the DPM projects.
Very grateful thanks go to Prof. Leonard Berliner from the New York Methodist Hos-
pital. Dear Lenny, thank you for your completely selfless support since we first met several
years ago. I always appreciated your interest in my research progress, discussions about
my work and challenging of my thoughts.
I also thank Dr. Agnieszka Oniśko from the Bialystok University of Technology, Prof.
Marek J. Druzdzel from both the Bialystok University of Technology and the University of
Pittsburgh, and Prof. Bernhard Preim from the University of Magdeburg. You supported
VIII Acknowledgements
me in the last steps of my research with incomparable open-minded and fruitful discus-
sions, thank you.
Finally, I would like to thank my dear friends, family and, especially, my wife Melanie.
You all gave me the encouragement to reach my aspirations. Dear Melanie, you enabled
me to follow the sense of life by developing knowledge and new life. Thank you for our
wonderful son Leonardo. Love you.
Contents
Page
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XV
List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . X
. VII
I Basic Consideration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
7 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.2 Prediction in a Clinical Treatment Decision Context . . . . . . . . . . . . . 69
7.3 Validation and Modification Workflow . . . . . . . . . . . . . . . . . . . . 70
7.4 TNM Model and Study Set-up . . . . . . . . . . . . . . . . . . . . . . . . 71
7.5 Quantitative Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.6 Qualitative Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.7 Results and Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.8 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.1 Direct dependencies between four variable types: personal data, anomalies,
examination data, and deterministic decisions. . . . . . . . . . . . . . . . . 60
6.2 TreLynCa: A BN model for treatment decision support of laryngeal cancer. 64
6.3 The TNM staging subnetwork from the treatment decision model of
laryngeal cancer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
9.1 An exemplary GUI for PSBN explorations, with a use case of TNM
verification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.2 GUI for PSBN exploration: Details of a node. . . . . . . . . . . . . . . . . 109
9.3 GUI for PSBN exploration: Examples of circle-charts comparing observed
and unobserved variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.4 GUI for PSBN exploration: Example glyphs for the comparison of two
TNM model computations. . . . . . . . . . . . . . . . . . . . . . . . . . . 111
9.5 Verification workflow of a TNM staging given in a patient record. . . . . . 112
9.6 Study results of the verification tool with five participants and 21 patient
cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
List of Tables
3.1 First table with recommendations and best practices for the development
and implementation of a CDSS, compiled from Berner’s book on CDSS. . . 24
3.2 Second table with recommendations and best practices for the development
and implementation of a CDSS, compiled from Berner’s book on CDSS. . . 25
3.3 Third table with recommendations and best practices for the development
and implementation of a CDSS, compiled from Berner’s book on CDSS. . . 26
8.1 Table of negative aspects from the CMT usability study . . . . . . . . . . . 101
9.1 Personal details (self-assessments) and results from the five participants of
the verification study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
List of Abbreviations
ICCAS Innovation Center Computer As- TDSS therapy decision support system
sisted Surgery TPU therapy planning unit
ICT information and communication
technology
ID influence diagram UID user interface design
List of Equations
Basic Consideration
Chapter 1
Introduction
Since the report To Err is Human of the Institute of Medicine in 2000 [88], computer-aided
clinical decisions are considered to be essential for a significant improvement of patient
safety. First diagnostic clinical decision support system (CDSS) were developed in the
1970’ with The Internist and MYCIN, and became more popular, widely accepted, and,
today, clinically desired since computers are faster, user interfaces are more intuitive and
systems provide high quality suggestions. When CDSSs support clinicians with the right
information at the right time they have the potential to ”[...] change the way medicine has
been taught and practiced” (Berner, 2009).
However, most CDSS projects fail to become clinically integrated because of an ini-
tially insufficient strategy regarding their development, clinical adoption and maintenance
as well as disregarded user requirements. Especially for complex treatment decisions, a
CDSS development and acceptance may be hampered due to the less verifiable nature of
decisions.
1.1 Motivation
This work is motivated and supported by the Digital Patient and Process Model (DP2M)
Project at the Innovation Center Computer Assisted Surgery (ICCAS) in Leipzig, Germany.
DP2M aims to develop a computer system that supports clinicians in finding the best ther-
apy decisions for patients with complex diseases. In case of complex diseases, patients
require individually tailored treatment decisions in order to reach the qualitatively best
possible outcomes or simply to survive. Therefore, multidisciplinary teams of clinicians
are required to collect, study and discuss extensive sets of patient data in order to find the
best treatment options by mutual consent. CDSSs have the potential to support difficult
decision-making tasks. For the basis of a CDSS, the Bayesian networks selected by DP2M
provide a transparent, human-understandable and customizable knowledge representation
and decision reasoning. Laryngeal cancer has been selected as the first exemplary disease,
and meetings of multidisciplinary expert teams as the starting therapy decision support
environment. This project collaborates closely with the Head and Neck Department at the
University Hospital in Leipzig, Germany.
1.2 Objective
The main objective of this dissertation is to contribute to the methodology of Clinical De-
cision Support System development. Specifically, I focus on building a Therapy Decision
Support System that is based on Bayesian networks. Contributions are mainly presented
in context of the support of multidisciplinary therapy decisions for patients with laryngeal
cancer.
A further aim of this thesis is to define a strategy that includes methods, systems and
tools as well as users involved in developing, validating, maintaining and interacting with
therapy decision models that are based on Bayesian networks. From this strategy, the focus
lies mainly in 1) the knowledge representation of complex multidisciplinary treatment de-
cisions, 2) the knowledge engineering for building these complex Bayesian networks when
adequate data is unavailable, and partially on 3) the visualization for user-model interac-
tion. The knowledge engineering of Bayesian networks includes graphical and probabilis-
tic modeling as well as model validation. A visualization is presented in context of an
expert-based treatment analysis.
1.3. Method 5
1.3 Method
A Therapy Decision Support System based on Bayesian networks is the given primary
objective. First, an analysis of the clinical workflows was conducted to identify situations
of decision-making, the participants and their environments.
A concept for a Therapy Decision Support System that is based on Bayesian networks
was defined. Further conceptions extended the existing Bayesian network methodologies
of building complex models, knowledge engineering, and graphical user interfaces. A
conceptualized model structure enables modeling multidisciplinary therapy decisions and
was applied to laryngeal cancer. A concept for knowledge engineering reduces the role of
knowledge engineers (usually computer scientists) by a clinically adopted modeling sys-
tem, which enables domain experts in autonomous modeling and validation. A visualiza-
tion concept focuses on the requirements in context of the users and environments.
Concepts were exemplary implemented and validated or evaluated. The model of laryn-
geal cancer was validated in teamwork between a knowledge engineer and a domain expert,
and using a dataset that is relatively small compared to the model size. From the knowledge
engineering concept, two tools for autonomous and collaborative modeling were imple-
mented and validated by clinicians; one for graph- and the second for probabilistic mod-
eling. Additionally, a usability study was conducted for the probabilistic modeling tool.
From the visualization concept, a visualization tool was developed to support analyses of
complex patient-specific models, and finally evaluated by a group of clinicians.
2.1.1 Epidemiology
Laryngeal cancer has a worldwide annual incidence of approximately 157,000 cases [56].
In Germany, the annual incidence reaches approximately 3,600 cases, which amounts to
25%-30% of all head and neck cancer cases [78, 136]. The five-year survival rate is about
47%. Over the last years, a slight increase of survival is observed through earlier diag-
noses and optimized integrated treatment approaches [22]. Since survival increases, more
attention is paid to patients’ quality of life and enhancement of functional outcomes [57].
Based on an in-house study from 2012 to 2015, the University Hospital Leipzig has
an annual incidence of approximately 60 new laryngeal cancer cases and nearly 20 recur-
rences.
Figure 2.1: Gray’s [65] sketches of the laryngeal area representing a) a larynx surface, b) a coronal
section of the larynx, and c) lymph nodes surrounding the larynx (highlighted by blue color).
and phonation (for pitch and volume). The most important role for survival is to protect the
lungs from foreign objects by coughing and other reflexive actions.
lymph nodes surrounding the larynx, and through the lymph channels or the bloodstream
to potentially all other organs. The original cancer is called primary tumor, while cancer
cells that migrated to other tissues and formed descendeant tumors are called metastases.
A primary tumor of laryngeal cancer can be located in all five laryngeal areas. First
metastases usually appear in locoregional lymph nodes. Distant metastases mainly oc-
cur in the lung, liver and bone through the bloodstream. Symptoms come from laryngeal
dysfunctions: problems with swallowing or change of the phonation as well as swollen
lymph nodes. For a patient, to provide accurate diagnostics, usually, a number of diagnos-
tic methods are evaluated, each with advantages and disadvantages. Diagnostic standard
examinations of the larynx area are performed by clinical examination, endoscopy, palpi-
tation, pan-endoscopy, ultrasound, medical imaging techniques (e.g., computer tomogra-
phy, magnetic resonance imaging and positron emission tomography) or histopathology.
Stroboscopy gives information about the vocal folds’ functionalities. Panendoscopy un-
der general anesthesia allows a direct view on the cancer surface, and is usually combined
with a biopsy (tissue removal). Medical imaging techniques give an insight about the or-
gan structures and metabolic processes, but findings may be difficult to differentiate from
other anomalies. Histopathology is mandatory for verification of the diagnosis from a tis-
sue biopsy, however, correct histological assessment depends on the available biopsy, that
is analyzed by the pathologist.
• N ∈ {N x, N 0, N 1, N 2, N 3}, and
• M ∈ {M 0, M 1}.
In general, the T-state relates to the number and type of infiltrated tissue, and vocal fold
functionality; the N-state relates to the number and localization of infiltrated nodes; and
the M-state with its two states describes the existence of metastases. In Appendix A, these
values are presented in more details.
The diagnostic evaluation and staging prior to treatment typically merges into a clinical
TNM staging (cTNM) and a (histo)pathological TNM staging (pTNM). A cTNM com-
bines data from physical examination, endoscopy and diagnostic imaging. A pTNM is
stated from (histo)pathological examination. In direct comparison between the cTNM and
10 Chapter 2. Tumor Board Decision for Larynx Cancer Patients
pTNM, the pTNM is considered to be more reliable. However, the correct diagnosis relies
on the histopathological report and consequently depends on the quality of the specimen.
Therefore, a correct TNM staging depends on the availability and correctness of the exam-
ination results and their individual quality.
Finally, for laryngeal cancer there exist three basic treatment options: surgery, radia-
tion therapy and chemo therapy. However, each treatment can be combined with the other
in various orders. The best combination of treatments depends on the individual patient
characteristics including the TNM staging, tumor resectability, comorbidities, general con-
dition, potential risk factors and quality of life, and in the near future also genetic factors.
Figure 2.2: A representation of EBM, which is based on three factors: 1) randomized controlled
trials, 2) clinical experiences, and 3) patient values and preferences.
In general, external clinical evidences are usually based on randomized controlled tri-
als from basic medical research. Researchers aim for identifying therapies that increase
the patient survival and quality of life, and decrease unnecessary procedures and costs
[86, 140]. Best evidences are collected in clinical protocols, called guidelines. Clinicians
are advised to follow regularly updated guidelines to provide the best available patient
care [165]. However, even excellent evidences are limited to 1) populations, 2) predefined
2.3. Head and Neck Tumor Board 11
sets of parameters, and 3) the number of studied patient characteristics [48]. Furthermore,
patient-specific values or preferences may disagree with guideline recommendations, but
they are crucial for any clinical decision [140]. Therefore, clinical knowledge and experi-
ence is essential to tailor guideline recommendations and preferences to a specific patient
case, also known as clinical judgment [140, 149].
Specifically to find the best treatment options for complex patient cases (e.g., with rare
or multiple diseases, or unusual health conditions), clinical judgment requires the exper-
tise of a multidisciplinary team. The increasing number of available examination methods
(e.g., physical examinations, imaging, and histology) and treatment options (e.g., surgeon,
radiation therapist and system therapist) theoretically allows for a patient-specific diagnosis
and an individually tailored treatment. However, an optimal EBM by a multidisciplinary
expert team may become impossible due to 1) the large amount of information with 2) a
less verifiable nature that is 3) mentally processed and, therefore, difficult to 4) communi-
cate, 5) understand, and 6) reproduce [55, 84]. In detail, complex patient cases may require
treatment combinations that are neither covered by guidelines nor well experienced by
clinicians. Clinical decisions (including deviations) are partially subjective and, therefore,
may be difficult for others to understand and reproduce [84]. For instance, a clear under-
standing of complex thought processes is necessary to be reproducible as regards insurance
providers’ management of costs and treatment errors. Furthermore, wrong decisions are
also costly by means of losing both the patients’ trust in health care systems as well as the
satisfaction of patients and clinicians [88]. Finally, it seems to be a paradox that clinicians
are advised to follow guidelines being responsible for any deviation, but complex patient
cases require deviations [55].
experts are from various domains, specialized in a subset of examination methods and
therapy procedures of usually one body area. Each method and procedure brings its own
complexity. A specialization in combination with an expert’s background knowledge and
experiences can be described as the expert’s viewpoint. Finally, the scope of a mental
patient model and, therefore, the quality of decisions depends on the clinicians’ ability to
gain and combine the other viewpoints.
Figure 2.3: A head and neck tumor board at University Hospital Leipzig with experts from
multiple disciplines. The complexity of decision-making is exemplary represented by a
visualization of decision models above the experts’ heads. Models vary depending on an expert’s
viewpoint, all representing the same patient.
2.3-A1 to share information among participating physicians to increase the quality of care
for patients,
Introduction
CDSSs aim for improving clinical decision-making and patient safety [88]. They can im-
prove patient care by ”providing the right information to the right person at the right point
in workflow in the right intervention format through the right channel” [126].
CDSSs may support decisions passively, e.g., by error recognitions using computer-
ized physician order entry (CPOE) and monitoring a patient situation, or actively, e.g., by
providing alerts in unusual or dangerous situations, recommending medications, and sup-
porting physicians to find optimal decisions. Active decision support systems that require
a rethinking and reorganizing of decisions and health care plans from clinicians (e.g., sup-
port for diagnostic and treatment decisions) are more likely to fail their acceptance [154].
In such cases, clinicians ignore or overwrite the decisions of a system and, finally, stop
using it.
Once a CDSS is built, its clinical acceptance depends on an appropriate integration, by
means of both its technical adaption to existing clinical systems, specifically to the local
electronic health record (EHR), as well as user-friendly interfaces. Well developed and
clinically integrated, CDSSs can minimize errors, promote patient safety, save time and,
finally, decrease the costs of care [88]. A successful CDSS development and clinical inte-
gration requires to reach one of these expected benefits without impairing the remaining,
or at least to increase the cost-benefit ratio [88]
The work of Berner et al. [12] reviews CDSSs to investigate their impact and effective-
ness on clinical decision-making and points out various challenges. Challenges concern
technical issues, such as data integration, system development, issues around the vocabu-
lary, system output and maintenance, as well as organizational and personal issues, such
as vendors, developers and users, and, finally, legal and ethical issues. To address all these