View
2
Download
0
Category
Preview:
Citation preview
Conformance Checking for Care Pathway ComplianceAssessment
Sebastião Salema Cordeiro Aboim de Barros
Thesis to obtain the Master of Science Degree in
Biomedical Engineering
Supervisors: Prof. Doutor Mário Jorge Costa Gaspar da Silva
Dra. Francisca Maria Pais Horta Leite
Examination Committee
Chairperson: Prof. Mónica Duarte Correia de OliveiraSupervisor: Prof. Doutor Mário Jorge Costa Gaspar da Silva
Member of the Committee: Jorge Manuel Graça Teixeira Santos
October 2018
Acknowledgements
I would like to thank everyone that supported me during the adventure of doing my Masters
and in the development of this dissertation. A special acknowledgement to those who never let
me go through this process by myself.
Firstly, I would like to express my sincere gratitude to Professor Mario Silva for the countless
hours of meetings and e-mails exchanged, as well as for his mentorship throughout the entire
process. Also, I would like to thank Francisca Leite and Nuno Silva, from Luz Saude, for their
constant follow-up and advice that guided me in the right direction.
I express my thanks to INESC-ID and to UpHill, for establishing the contact with Luz Saude
and making this project possible.
I would also like to express my deep gratitude to my girlfriend, Mariana Costa. This work
would not have been possible without your support and motivation. Thank you for your patience
and for helping not to doubt myself.
To my family for supporting me in every possible way and for encouraging me to never
give up and to achieve the best results in life. Thank you for being the example of effort and
devotion, always granting me the best opportunities.
Finally, I would like to thank all my friends, specially to David Honorio and Eduardo
Rodrigues, for helping me through every stage of this work. To their companionship and support
that helped me grow as a person, who are always there in the good and bad times, day after
day.
iii
Resumo
A necessidade de prestar tratamento medico de excelencia e servicos de saude de alta quali-
dade tem vindo a aumentar continuamente, o que leva a crescente adocao de protocolos clınicos
por organizacoes de saude. Contudo, tem-se observado que a pratica clınica se desvia das re-
comendacoes clınicas, o que se pode refletir em nıveis inferiores de qualidade do servico e dos
resultados clınicos. Neste sentido, monitorizar a conformidade da realidade face ao percurso
clınico torna-se essencial para a otimizacao dos processos de saude.
O presente estudo define um modelo que permite avaliar a conformidade entre a pratica
clınica observada e um percurso clınico, com recurso a tecnicas de prospeccao de processos
suportado por metricas definidas com base em cenarios de nao conformidade, considerando
atividades nao efetuadas e inseridas. O modelo proposto examina 345 casos clınicos do Hospital
da Luz Lisboa com Cancro Colorretal para os quais a conformidade foi estudada, proporcionando
uma analise detalhada sobre as intervencoes executadas.
Os resultados obtidos permitiram concluir que nenhum caso clınico contem dados que evi-
denciem conformidade total com o protocolo. Em media, apenas metade do percurso clınico foi
completado, o que mostra que existe ainda espaco para melhoria. Quanto aos resultados clınicos,
verificou-se um numero inferior de dias pos-operatorio para casos com maiores nıveis de con-
formidade (IC=95%). Desta forma, este trabalho oferece um modelo robusto para monitorizar
a pratica clınica com o objetivo de melhorar a qualidade dos cuidados de saude.
v
Abstract
With the growing adoption of care pathways worldwide, the urge to deliver high-quality
healthcare services and treatment excellence is constantly increasing. However, a gap has been
observed where clinical practice often deviates from the clinical best-practices model, which
can lead to lower levels of quality care and patient outcomes. It is essential for healthcare
improvement to monitor and fully understand how reality conforms to the care pathway.
Built on process mining techniques, supported by a set of metrics computed on non-
compliant scenarios considering skipped and inserted activities, this study devises a model to
assess the compliance between the clinical practice and the specifications in a care pathway.
The proposed model addressed 345 clinical cases from Hospital da Luz Lisboa with Colorectal
Cancer, divided into colon and rectum cancer patients cases, and reviews the conformance level
for each group, providing diagnostic compliance analysis on the medical behaviours.
Results on the real-world dataset allowed to conclude that no patient case had data evi-
dencing full compliance with the pathway. On average, only about half of the pathway was
completed, proving that there is room for improvement. Concerning clinical outcomes, results
showed that the number of post-operative days until discharge was lower for patient cases with
higher compliance levels (CI=95%). Therefore, this work provides a robust model for monitoring
clinical practice as an attempt to increase the quality of healthcare.
vii
Palavras Chave
Keywords
Palavras-Chave
Percurso Clınico
Analise de Conformidade
Prospeccao de Processos em Saude
Verificacao da Conformidade
Medida de Conformidade
Cancro Colorrectal
Keywords
Care Pathway
Compliance Analysis
Process Mining in Healthcare
Conformance Checking
Fitness
Colorectal Cancer
ix
Contents
1 Introduction 1
1.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Structure of the Document . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Concepts and Related Work 7
2.1 Understanding Care Pathways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Healthcare Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Care Pathways and Clinical Guidelines . . . . . . . . . . . . . . . . . . . . 8
2.1.3 Adopting Care Pathways . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.4 Challenges of Implementing Care Pathways . . . . . . . . . . . . . . . . . 10
2.1.5 Compliance Checking and Monitoring . . . . . . . . . . . . . . . . . . . . 11
2.2 Modeling Care Pathways with BPMN . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Process Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.1 Concepts of Process Mining . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 Event Logs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.3 Process Mining Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.4 ProM Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Evaluating Compliance to a Care Pathway . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
xi
3 Care Pathway Compliance Assessment 23
3.1 Architecture of the Model Based on Non-compliance Criteria . . . . . . . . . . . 23
3.2 Conformance Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Metrics to Examine the Compliance of Patient Cases . . . . . . . . . . . . . . . . 29
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4 Colorectal Cancer Care Pathway Compliance Analysis 33
4.1 Dataset Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Overall Compliance Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Compliance Evaluation and Interpretation . . . . . . . . . . . . . . . . . . . . . . 37
4.3.1 Patients Distribution by Compliance Levels . . . . . . . . . . . . . . . . . 37
4.3.2 Compliance for Each Activity of the Pathway . . . . . . . . . . . . . . . . 38
4.3.3 Compliance by the Phase of the Pathway . . . . . . . . . . . . . . . . . . 39
4.3.4 Compliance of the Inserted Events of the Pathway . . . . . . . . . . . . . 40
4.3.5 Clinical Outcomes Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5 Conclusions and Future Work 47
5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Bibliography 54
A ProM – Conformance Checking 55
xii
List of Figures
1.1 Structure of the dissertation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Example of the aneurysm diagnosis pathway. . . . . . . . . . . . . . . . . . . . . 9
2.2 BPMN symbols used for the graphical representation of a process model. . . . . 13
2.3 BPMN model of the care pathway from Figure 2.1. . . . . . . . . . . . . . . . . . 14
2.4 Process mining techniques in healthcare. . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 The ProM framework. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1 Overview of the proposed model. . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Architecture of the proposed model to assess the compliance to a care pathway. . 25
3.3 Missing activities of a trace. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4 Inserted events of a trace. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1 Distribution of patient cases by the number of events occurred. . . . . . . . . . . 36
4.2 Distribution of the compliance values for each patient case. . . . . . . . . . . . . 38
4.3 Distribution of all patient cases by level of compliance. . . . . . . . . . . . . . . . 39
4.4 Average post-operative days until hospital discharge per compliance group (Colon) 42
4.5 Average post-operative days until hospital discharge per compliance group (Rectum) 43
A.1 Conformance checking between the rectum patients careflow and the pathway. . . 55
A.2 Conformance checking between the colon patients careflow and the pathway. . . . 56
xiii
xiv
List of Tables
2.1 Example of an event log . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Overview of CP compliance analysis studies. . . . . . . . . . . . . . . . . . . . . . 20
3.1 Conformance measurement for the pathway modelled in Figure 2.3. . . . . . . . . 29
3.2 Metrics computed to support the compliance analysis. . . . . . . . . . . . . . . . 30
4.1 Statistical characterization of the colorectal cancer dataset. . . . . . . . . . . . . 34
4.2 Proportion of patient cases arranged by the number of events occurred. . . . . . 36
4.3 Overall alignment results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4 Overall compliance results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.5 Frequency of occurrence for each activity of the CRC care pathway. . . . . . . . 40
4.6 Average frequency of occurrence for the phases of the colorectal cancer pathway. 40
4.7 Compliance results on the alien events observed for the colon cancer patients. . . 41
4.8 Average compliance on the post-operative days until hospital discharge. . . . . . 42
4.9 Kruskal-Wallis results on the post-operative days until hospital discharge. . . . . 43
4.10 Average compliance on the mortality. . . . . . . . . . . . . . . . . . . . . . . . . . 44
xv
xvi
Abbreviations
CRC Colorectal Cancer
CP Care Pathway
HCP Health Care Professional
EHR Electronic Health Record
KPI Key Performance Indicator
EBM Evidence-based Medicine
LOS Length of Stay
BPMN Business Process Model Notation
BPM Business Process Management
EPC Event-driven Process Chain
HIS Health Information System
MDA Multidisciplinary Appointment
CEA Carcinoembryonic Antigen
xvii
xviii
1Introduction
Due to the rapid development of hospital information systems, Electronic Health Records
(EHR) are becoming more complex and complete, to cope with the large series of patient data
being captured. Simultaneously, science is continuously evolving and advanced technology is re-
leased every day, increasing the need to deliver high quality healthcare services (van de Klundert
et al., 2010). In such systems, healthcare professionals (HCP) record their medication and ex-
amination prescriptions, appointments notes, diagnostic hypothesis, etc. This opened the door
to mining clinical data to make the clinical practice better for the care of individual patients.
The collected data contains details on the patients’ progression, which is along or around the ex-
pected course expressed in the care pathways (CP). This provides the opportunity for studying
medical cases, clinical outcomes and to extract relevant knowledge about the medical treatment
behaviours (Yan et al., 2017).
In many domains, such as manufacturing and services industries, when a process is exe-
cuted, there are observed discrepancies between the expected and the actual behaviour and,
consequently, in the outcomes. Medical services are no exception and the reality often deviates
from the clinical practice guidelines (Rovani et al., 2015). Thus, this gap between the clinical
recommendations and the clinical practice is a matter of utmost importance in healthcare man-
agement, which is shown, for instance, in Hajjaj et al. (2010) for dermatology, or in Cornu et al.
(2010) for urology. Moreover, according to the British Medical Journal1, the cost of medical
error is increasing to more than 17 billion dollars annually (van den Bos et al., 2011).
Care pathways provide detailed guidance for each stage in the management of a patient
(treatments and interventions), but multiple variances may still unavoidably occur due to in-
dividual complexities and subjectivity (Yan et al., 2018). As independent decision-makers,
personal experiences, background and logistic are some of the aspects why HCPs do not always
comply with the defined process model (Rovani et al., 2015). Some of these deviations may
benefit the patient; however, considering CPs as the treatment recommendation to be followed,
1https://www.bmj.com/
1
based on the premise of increasing the quality of care, they frequently represent flaws that can
jeopardize the patient’s health (Every et al., 2000).
Although the need for process monitoring is indispensable, the administration cannot man-
ually monitor all activities performed by physicians for all patient cases and locate violations
without allocating multiple resources (Quaglini et al., 2001). However, by comparing the EHR
collected data with the prevailing CPs, we can identify where the care given has deviated from
the care expected and quantify the compliance level (Yan et al., 2017). Compliance monitoring
is a crucial step to be considered when an organization intends to accomplish a higher degree
of process maturity (Weidlich et al., 2011), specially in healthcare organizations, due to the
complexity of medical treatment processes.
In this context, the need for compliance checking is becoming increasingly relevant to lever-
age a cost-efficiency in the quality of the care provided, to reduce unnecessary variations through
standardization of treatment behaviours and to improve CPs. Once non-compliance cases are
detected, the healthcare organizations can update their CPs or adopt new techniques to enforce
the best practice executions (Weidlich et al., 2011).
At Hospital da Luz Lisboa there is an effort to implement and define care pathways and
business intelligence. I was offered the opportunity to study how to assess compliance of medical
behaviours to an implemented CP and analyze the clinical outcomes with colorectal cancer
(CRC) patient data.
1.1 Objectives
Based on the premise that care pathways are used to increase the healthcare quality, the main
goal of this dissertation is to provide a model with a set of metrics that allows to measure the
compliance of the clinical executions to the expected course of care. The information collected
enables healthcare organizations to efficiently understand how HCPs are operating and audit
for deviation. Therefore, the idea is not to determine whether or when adherence to a CP is
justified, but to address whether medical interventions are compliant to a care pathway.
In brief, the objectives set for this dissertation were:
1. Review conformance assessment methods used for evaluating compliance to a process
model, in particular, to care pathways.
2. Devise a model to evaluate the compliance of clinical practice with the specifications in
2
a care pathway. For this purpose, identify an algorithm to measure conformity of HCPs
to the pathway for individual patient cases and provide a detailed compliance analysis on
the treatment behaviours.
3. Validate the efficiency of the proposed model in a real-life case study. Establish an asso-
ciation between the compliance levels achieved for the patients careflow with the clinical
outcomes of the patients under study.
1.2 Contributions
This dissertation, in line with the motivations and objectives set, offers the strategy to
effectively examine the performance of HCPs according to a pathway, enabling the process
improvement, the detection of predominant deviations and to understand how well adjusted to
the reality the clinical recommendations in a CP are.
This work defines a model to assess the compliance between the clinical practice and a care
pathway, considering skipped activities (activities that should occur according to the CP and
were not observed in clinical practice) and inserted activities (activities that occur but are extra
to the CP) of the pathway model. It measures to what extent the clinical executions conform
with the CP through a fitness metric computed with the algorithm proposed by Adriansyah et al.
(2011). The model provides additional computed clinical domain-specific metrics that allow to
collect insights on the compliance of individual activities, the order and pathway constraints,
for the inspection of the underlying non-compliant events. The proposed model enables to
gather compliance analytics on the medical behaviours and to segment a population of cases
into particular characteristics. It provides a proper monitoring of the clinical practice that can
work as an internal auditing system and high-quality assurance for an healthcare organization.
The second main contribution of this dissertation is the compliance analysis of a real-world
dataset of colorectal cancer patients, provided by Hospital da Luz Lisboa. It allowed to report
the reality of medical diagnostic and treatment interventions for this disease, giving an overview
of the non-compliant events observed, and to validate the proposed model. Also, to establish
an association between the compliance levels achieved and the impact in the patients’ health.
Results showed that the model is robust and feasible, reporting valuable information on the
compliance of HCPs with the hospital’s CRC pathway.
3
1.3 Methodology
In the first stage of the work, I studied the patient careflow of the colorectal cancer (CRC).
After Hospital da luz Lisboa granted access to the pathway and key performance indicators
(KPI), I performed a hands-on observation on the different steps associated with the diagnosis,
staging and treatment phases of the CP. This included the study of the activities and sequence
of actions that an HCP should execute and, therefore, should be observed in the data collected.
Hospital da Luz Lisboa runs private training sessions in a simulated environment on this subject
and I had the chance to attend several meetings where CRC was unceasingly discussed, and the
compliance issue was raised. This comprehension on the CRC CP played an import role in
establishing my awareness to recognize that the lack of compliance is a real problem.
The second stage consisted in the literature review. The course of this research followed two
questions that needed to be addressed:
1. What is the importance of compliance checking to a care pathway?
2. How is it measured and what insights can we gather with it?
The study started on the care pathway and clinical guidelines concepts and applicability, followed
by an extensive understanding of the importance of compliance checking for healthcare providers.
It was followed by the study of related work on similar problems on how to assess compliance
to a process model. The research took special consideration to the available algorithms for
measuring deviations to a given process. Even though there are several reports presenting very
interesting methods, they require high-quality data or exclusively examine CPs from an external
perspective, considering only the clinical outcomes (for instance, length of stay or infection
rates). The topic of data mining techniques – process mining – is gaining increasing attention
in healthcare management, as they allow to extract non-trivial knowledge about healthcare
processes. Therefore, the research focused on compliance assessment specifically using process
mining techniques of conformance checking. I selected an algorithm to measure the compliance
for each patient case and ideas from various of these publications were taken into consideration
and integrated in the proposed model.
The third stage involved the analysis of the CRC dataset obtained from the hospital’s clinical
system. After receiving the CRC dataset, the first step was (1) to identify the available activities
for each phase of the CP; (2) to map those activities in the clinical process; and (3) to identify
missing data. The dataset reflects information about the KPIs the hospital is collecting on the
4
CRC CP and does not include all treatment behaviours. Thus, only the activities of the pathway
matching available events recorded in the dataset were considered. Based on these activities,
two pathways were modelled, one for colon subset of patients and another for the rectum ones,
with the use of Business Process Management Notation (BPMN) modelling language.
The fourth stage of this work was the data preparation (extraction and transformation) of
the CRC data from Hospital da Luz Lisboa. The dataset was split into two subsets, one for the
colon cancer patient cases and another for rectum cancer patients. The data was converted from
a business intelligence dataset into a standard format (event log) containing all the activities a
patient was subjected to, representing the patients careflow. In such format, each entry is an
activity that occurred for a case in a specific timestamp. This was followed by the implemen-
tation of the proposed model and by the definition of the compliance analysis metrics based on
the possible non-compliant scenarios identified. By cause of its popularity and extensive public
documentation, Python was the selected programming language to develop this work.
The fifth and final stage of this work consisted in achieving the results for the compliance
assessment between the pre-processed dataset and the CRC CP. I performed a compliance
analysis, reporting on the overall compliance levels, critical activities and main violations, and
statistical tests were conducted to establish associations between the compliance results achieved
and the clinical outcomes recorded. The results were presented to the management team of
Hospital da Luz Lisboa, in order to provide them the information collected on the performance
of HCPs in the treatment of CRC.
1.4 Structure of the Document
This dissertation follows the structure illustrated in Figure 1.1. The next chapter, Chapter
2, introduces the concepts and related work including a description of healthcare processes, the
characterization of guidelines and care pathways and details on the importance of compliance
checking and its underlying analysis. Furthermore, process mining concepts are depicted in this
chapter in combination with previous work on compliance evaluation.
Chapter 3 describes the proposed CP compliance assessment model, involving the confor-
mance measurement for each patient case in line with the computed metrics to support the
compliance analysis.
Chapter 4 presents the results from the CRC CP compliance assessment. It starts with a
brief statistical characterization of the dataset. It focus on the overall compliance levels achieved
5
Figure 1.1: Structure of the dissertation.
and on the insights gathered for the compliance to the pathway, as well as on the associations
with the clinical outcomes.
Finally, Chapter 5 summarizes the key contributions of this work and its limitations, and
discusses the possibilities for further developments in the future.
6
2Concepts and Related Work
The conformance between the treatment behaviours and the clinical recommendations is
recognized as a complex issue, for which the role of compliance monitoring applied to care path-
ways (CP) is identified as a topic that must receive extreme attention. This chapter describes
fundamental concepts on healthcare processes, previous studies and techniques addressing the
measurement of compliance of process executions to a process model. Section 2.1 presents the
concepts needed to understand care pathways, introducing their benefits and challenges, as well
as the importance of compliance assessment. Section 2.2 describes the modelling of CPs with
the process modelling language Business Process Management Notation1 (BPMN). Section 2.3
details data mining concepts and techniques for process execution data – process mining. Sec-
tion 2.4 reviews previous work in the field of compliance and deviation measurement, with focus
on CPs, including process mining techniques applied to the healthcare environment. Finally,
Section 2.5 concludes with an overview of the state-of-the-art.
2.1 Understanding Care Pathways
2.1.1 Healthcare Processes
Healthcare processes are commonly viewed as operational business processes for healthcare
organizations. These processes are related to the set of clinical and non-clinical activities or
operations, within a particular healthcare organization, in a specific structure and sequence,
that co-operate to diagnose, treat and prevent any disease, aiming at improving the patients’
health (Rojas et al., 2016). Unlike industry processes, healthcare processes often bear no relation
to the ideal and expected scenarios (Huang et al., 2013). In fact, the healthcare environment
and its underlying processes, which are continuously changing and evolving, are described in
the literature as highly dynamic and complex, ad-hoc and multi-disciplinary-based (Rebuge and
1http://www.bpmn.org/
7
Ferreira, 2012).
Healthcare processes can be classified into: medical treatment processes and generic organi-
zational processes (Rojas et al., 2016). The latter ones, also known as administrative processes,
are “generally handled according to well-established procedures”, which support and coordinate
inter-operating medical treatment processes for healthcare professionals (HCP) and units (Lenz
and Reichert, 2007). Examples of such processes are patient scheduling, medical order entry or
results reporting.
On the other hand, medical treatment processes, or clinical processes, are patient-focused
and concern about diagnostic and therapeutic actions to treat the patient, which must be
adaptable to the variability of the healthcare environment. They are generally assigned to the
diagnostic-therapeutic cycle, which involves observation, reasoning, and action activities (Lenz
and Reichert, 2007). Medical treatment processes are executed in accordance with international
approved clinical recommendations that have proven to be effective (Rovani et al., 2015).
2.1.2 Care Pathways and Clinical Guidelines
In clinical practice, the patients careflows are usually subjected to recommended and stan-
dardized treatment interventions. A care pathway, also known as a clinical pathway, care maps
or critical pathway, is the term assigned to the evidence-based medical treatment processes of
operational nature that HCPs should act in accordance with.
CPs may have distinct meanings, depending on the stakeholders, such that interpretations
of a CP range from simple medication usage to the entire careflow of a patient. However, Huang
et al. (2012) defines a care pathway as:
A clinical pathway, guided by evidence-based medicine (EBM) and clinical prac-
tice guidelines, is a standardized and normalized therapy pattern and procedure con-
structed for a specific disease that follows the contemporary medical empirical data
and clinical experts’ experiences.
In addition, according to Every et al. (2000), a CP corresponds to a set of integrated
management plans to reach the patients’ health goals, which provide the course, sequence, and
timing of actions. Moreover, a CP is a multidisciplinary patient care plan, in which every
intervention that should be performed by medical professionals is well-defined and optimized
(Yang and Su, 2014).
8
Figure 2.1: Example of the set of activities to be performed for the aneurysm diagnosis pathway(van de Klundert et al., 2010).
CPs are viewed as algorithms, as exemplified in Figure 2.1, detailing structured processes
for the appropriate healthcare for a specific and concrete clinical setting and individual circum-
stances, designed by each organization (Basse et al., 2000). There are four main components that
can be distinguished in a CP: the timeline, categories of care and the interventions, outcomes
criteria and variance record (records on deviations to be documented and analyzed).
As expressed by Huang et al. (2012), care pathways are anchored in international approved
clinical recommendations, named clinical practice guidelines. The MeSH1 dictionary defines
medical guidelines as:
Set of directions or principles to assist the healthcare practitioner with patient care
decisions about appropriate diagnostic, therapeutic, or other clinical procedures for
specific clinical circumstances.
In contrast to a CP, guidelines are not site-specific and represent a set of consensus decla-
rations systematically developed to assist physicians in making patient management decisions
(Every et al., 2000). They aim to provide evidence-based and economically reasonable med-
1MeSH: Medical Subject Headings. Thesaurus and controlled vocabulary used for indexing the MEDLINE
database of medical literature provided by the National Library of Medicine.
9
ical treatment processes, to improve the patients outcomes and to decrease the variation of
healthcare quality.
2.1.3 Adopting Care Pathways
Implementing care pathways plays an important role in the clinical environment. CPs
are widely adopted by healthcare providers to deliver high-quality standardized care services,
in order to reduce unnecessary clinical practice variations, and to enhance the efficiency of
the medical treatment processes. Macario et al. (1998) corroborated that guidelines and care
pathways have a positive impact on the hospitals’ costs, proving a decrease in the hospital’s
average costs in a knee surgery after implementing the corresponding CP. Also, Marrie et al.
(2000) showed that the efficiency has increased with no harm to the well-being of the patients, in
a study of a specific pathway of pneumonia. Therefore, CPs improve both quality and efficiency
simultaneously, or one without influencing the other.
Regarding the patient clinical outcomes, Panella et al. (2003) showed that adopting CPs
lead to a decrease of 12% in patient mortality. Additionally, Quaglini et al. (2018) showed that
not only treating patients according to guidelines results in lower costs, but CPs also enable to
reduce the length of stay (LOS) and treatment complications, leading to a significant difference in
resources consumption. Such clinical recommendations allow to enhance resource management,
waste reduction, and support risk management, once they are expected to reduce clinical risk.
CPs also promote a better multidisciplinary communication, optimizing the care provided across
different departments and care planning between multiple HCPs (Yang and Su, 2014).
Therefore, implementing care pathways translates evidence-based medicine into clinical
practice and identifies which bad behaviours and malpractices must be removed. Nevertheless,
CPs are not prescriptive, since they do not override clinical judgment.
2.1.4 Challenges of Implementing Care Pathways
Currently, even though CPs are proven to be beneficial in healthcare (Weiland, 1997), the
introduction of CPs in an healthcare organization is a complex task. Generally, there are some
skepticism and difficulties in the adherence to CPs (van de Klundert et al., 2010). Additionally, it
takes time to have them accepted and used in the clinical practice, but also to adjust them to the
reality. In fact, it has been observed a gap between the treatment behaviours and the clinical
recommendations in a CP, where clinical practice often deviates from the pathways (Rovani
10
et al., 2015), as well as between the guidelines and the knowledge required to implement them
(Lenz and Reichert, 2007). Furthermore, the process of designing clinical pathways requires
a specific team (clinical experts, process participants and managers) committed to develop a
pathway for every disease according to the reality of the internal organizational structure. Care
pathways are often defined by large documents created in complex iterative processes and are
affected by the subjectivity of the creative team members and by the difficulties in translating
research findings into practice (Yang and Su, 2014).
Care pathways usually address medical treatment processes for the ideal patient and hospital
and may not adequately account for relevant issues to the everyday’s patient care (Rovani et al.,
2015). In fact, not all variation in a patient care is negative (Every et al., 2000). Each patient is
different and CPs generally deal with standard conditions better than unpredictable ones (do not
respond well to unexpected changes). For instance, they generally do not contemplate variables
like allergies or the coexistence of pathologies. As a result, CPs are not always adjusted to the
reality of clinical practice.
2.1.5 Compliance Checking and Monitoring
Due to the gap observed and the growing adoption of care pathways worldwide, clinical
pathway compliance checking is of utmost importance in the healthcare environment. As de-
scribed by Huang et al. (2012), care pathway compliance analysis is the process of (1) discovering
knowledge about how clinical activities impact on the patients careflow and whether they act in
accordance with a CP and (2) using the discovered knowledge for:
• Care pathway redesign and iteration of internal medical processes;
• Care pathway optimization;
• Clinical decision support;
• Medical deviation detection;
• Business management.
In practice, hospitals have information systems to capture the treatment behaviours and the
details of a patient treatment journey, along or around the CPs. By comparing such data with
the regulations of a CP, it is possible to assess the compliance of HCPs to the expected course of
care and understand the main deviations. Actually, if not properly adopted or adhered to, the
implementation of a pathway will probably be unsuccessful in enhancing the quality of medical
treatment processes (van de Klundert et al., 2010). As claimed by Huang et al. (2013), healthcare
organizations that already perform conformance checking usually have an incorrect, not updated
11
and oversimplified incorrect view of the actual conduct in a patient careflow. Healthcare orga-
nizations generally examine compliance to a CP only from an external observation, focusing
mainly on the clinical outcomes (e.g. length of stay, costs, infection rate, etc.). Although these
measurements provide valuable information, the diagnostic and therapeutic actions must be con-
sidered as well (Huang et al., 2016). Noting that, while some clinical cases are valid executions
of the CP and completely comply with it, others capture a certain share of the recommended
behaviour, reason why it necessary to measure compliance a posteriori.
Therefore, CP compliance and variance analysis is fundamental in every healthcare orga-
nizations to promote process maturity and process improvement, aiming to monitor and fully
understand how the reality conforms to the pathway, for an effective performance assessment.
It enables to identify education and training needs, to make patients safer and achieve better
outcomes. In fact, compliance to the international approved guidelines and best practices is
essential to reduce litigation costs, make a shift towards value-based high-quality healthcare and
to allow the standardization of care treatments. In case of non-compliant behaviours, “health-
care organizations can update their CP specifications to cover the respective case, or they can
impose new mechanisms to enforce the best clinical practices” (Huang et al., 2014).
2.2 Modeling Care Pathways with BPMN
Diagrams and charts are commonly used to explain and describe a process and the problems
it implies. Throughout this dissertation, Business Process Model Notation (BPMN) was chosen
to represent care pathways, as they can be defined in a flow chart, with all the decisions,
characteristics and sequence interventions they involve, including multidisciplinary processes.
BPMN (2.0) semantics has been recently formalized in order to be suitable for compliance
analysis (Gorp and Dijkman, 2013). Researchers have proposed using this notation to address
the complexity of CPs, considering it is standard and easy-understandable. It supports the
complex formulations of clinical scenarios, such as loose sequential constraints, multiple choices
and nested processes (Yan et al., 2018).
BPMN was chosen over other process modelling languages, like Petri Nets, due to the fact
that it is intuitive and it is seen as the industry standard for process modelling. A Petri Net is a
dynamic structure composed by transitions, representing the activity to be performed, a set of
places, containing one or more tokens, and a set of directed arcs connecting places and transitions
(Rozinat and van der Aalst, 2008). A transition is enabled when all the places connected to the
transition contain a token. When a transition is enabled, it fires and consumes one token from
12
Figure 2.2: BPMN symbols used for the graphical representation of a process model.
the places connected to it, producing a token to the all of its output places. This firing changes
the state of the process, characterized by the distribution of tokens.
The sequence flow logic of a process represented with BPMN can also be explained with
tokens. Firstly, a token is produced in the start event and, as the process evolves, every time a
token reaches an activity, the respective task is performed. After is it completed, then the token
is released to the outgoing sequence flow until it reaches the end event. Therefore, a BPMN
diagram just requires that there is always a true condition through a possible path, where every
scenario is covered and the process always ends.
The most common BPMN symbols used to model CPs are depicted in the Figure 2.2. BPMN
provides a symbolic notation for events, activities, gateways and relations. Events represent
events (“something that happens”) in a process: the start event acts as the process trigger; the
end event signals that the process has ended; and intermediate events indicates an event that
occurs between the start and end event. Tasks, represented by boxes, are the units of work to
be performed (“something that is done”) and, when containing a box with a “+” they refer to a
sub-process. Gateways are always represented by diamonds. The exclusive gateway, represented
by a diamond with an “X” (sometimes, it is simply the diamond), based on one condition,
breaks the flow into exactly one path from the set of paths. On the other hand, the diamond
with an “+” is the parallel gateway and does not depend on conditions or events. It is used to
illustrate that all paths are activated simultaneously and, thus, all paths must be completed.
The inclusive gateway, a diamond with a circle, breaks the process flow into one or more paths
that must be completed, for example, with this gateway, 2 out of 3 paths can occur, depending
on the conditions. Lastly, for relations, the sequence flow arrow defines the execution order of
activities and the association symbol represents a relationship between artifacts and objects.
An example of a care pathway modelled with BPMN is illustrated in Figure 2.3, using the CP
from Figure 2.1.
13
Figure 2.3: BPMN model of the care pathway from Figure 2.1.
Figure 2.4: Overview of the process mining techniques in healthcare (Rojas et al., 2016).
2.3 Process Mining
2.3.1 Concepts of Process Mining
Process mining, a method for business process management (BPM) and analysis, is gaining
increasing attention in healthcare (Mans et al., 2013). Business Process Management (BPM)
aims at supporting business processes using techniques to design, enact, control, and analyze
operational processes (van der Aalst et al., 2003). Process mining is an emerging discipline that
has been adopted in the domain of BPM, applicable to complement and support the design and
diagnosis phases, but also to restructure the life cycle itself (Huang et al., 2013).
Process mining aims at extracting knowledge and non-trivial useful information from the
process executions, to discover, monitor and improve processes (Mans et al., 2013). It pro-
vides a comprehensive sets of tools to return fact-based insights and, thus, to support process
14
improvement.
There are three main current types of process mining: discovery, conformance and extension
(Rojas et al., 2016) (see Figure 2.4).
1. Discovery is related to inferring process models that are able to reproduce the observed
behaviour. The outputed model can be represented with BPMN or another notations (e.g.
Petri Net or Event-driven Process Chain (EPC)).
2. Conformance checking allows to compare an a priori model with the observed behaviour,
to analyze if the reality conforms with the process model.
3. Extension/Enhancement corresponds to the projection of the information extracted form
the observed behaviour onto the model, where the goal is to enrich the model with the
data in the event log.
Despite the difficulties due to the richer diversity and complexity of medical behaviours
over other business processes (Huang et al., 2013), discovering process models and analyzing
their performance provides relevant opportunities for extracting knowledge out of the abundant
information recorded in an hospital’s EHR. Process mining ensures that not only the health-
care processes are well-understood, but also allows to achieve a high level of process efficiency.
With the usage of process mining techniques, healthcare organizations can discover the pro-
cesses as they are executed in reality, check whether certain practices were really followed and
gain insights into resource utilization and performance-related aspects. Thus, process mining’s
wide applicability in healthcare can help to construct or redesign the processes, to analyze and
improve the performance and collaboration between physicians, to identify which activities are
bottlenecks and to add alternative or supplementary data to the activities of the process (Rojas
et al., 2016).
2.3.2 Event Logs
The data captured in EHRs on the medical behaviours produce large event logs. These can
be viewed as lists, as exemplified in Table 2.1. An event log contains a set of traces, in which
each trace comprises all the activities executed for a singular process instance/case (Rojas et al.,
2016). Therefore, an event log consists of all the occurrences used to construct a journey map
(Neira et al., 2017). It is the input and raw material for running process mining algorithms, and
can have multiple attributes, apart from minimum structured required – the case and activities
identifiers.
15
Table 2.1: Example of an event log
CaseID Activity Originator Timestamp
4 X-Ray John 09-03-2018 15:01
2 X-Ray John 09-03-2018 15:23
3 Blood Test Michael 09-03-2018 16:01
3 X-Ray Susan 09-03-2018 16:45
4 Blood Test John 09-03-2018 17:24
1 Blood Test Susan 09-03-2018 18:37
1 X-Ray Michael 10-03-2018 09:11
2 Blood Test Michael 10-03-2018 10:13
3 Surgery John 10-03-2018 12:41
4 Discharged Michael 10-03-2018 16:29
3 Blood Test Michael 10-03-2018 19:06
2 Surgery John 11-03-2018 08:53
1 Surgery Susan 11-03-2018 10:18
1 Blood Test John 11-03-2018 11:39
Event logs can be classified on the completeness of the information and on the degree of
“noise” (Dongen and Frans, 2007). A log is considered complete when it contains all the possible
process behaviours. The noise refers to incorrect, exception or missing parts in the event (for
instance, skipping an activity).
Extracting the observed events from an information system and then relating them to the
activities of the process model may be a complex and delayed task (Weidlich et al., 2011).
However, to be able to apply the data to process mining techniques, the event log has to fulfill
some requirements (van der Aalst et al., 2007) (Hornix, 2007). Even though they can be different
in nature, we assume that, in the events that are recorded:
1. Each event refers to one activity, a well-defined step in the process;
2. Each event refers to only one case;
3. Each event has only one performer, also named originator;
4. Each event has a timestamp;
5. All events are ordered by time.
2.3.3 Process Mining Perspectives
Process mining techniques comprise not only the automatic discovering and structuring of
the process in different kinds of models, but also analysis techniques as an attempt to check and
analyze models and/or event logs, having the process mined/modelled a priori (Yang and Su,
16
2014). It is possible to distinguish three different perspectives of process mining (van der Aalst
et al., 2007): process perspective (the “How?”), organizational perspective (the “Who?”) and
case perspective (the “What?”).
The process perspective, also known as the control-flow perspective, attempts to discover the
control-flow – the actual order of the activities – which is, most of the times, the starting
point to determine the other perspectives. This perspective aims to define an acceptable
characterization of all possible paths to model/construct the process, from a complete
event log, comprising all the activities in the process and their relations (for instance with
a Petri Net).
The organizational perspective targets the “who?”, the performers of each activity (in the event
log) involved in the process and what are their relations. As a result, the relations between
the performers can be identified and classified according to the roles they play.
The case perspective’s purpose is to examine the properties of the cases, which can be described
by their paths in the whole process or by the performers involved in each case, but also
by the values of the respective data elements. With such information, one can discover
correlations and decision rules used to find other possible paths in the process.
Additionally, with respect to analysis techniques, process mining can also be applied for
performance analysis and conformance checking. In particular, the latter one, considered the
fourth perspective of process mining (Rojas et al., 2016), is employed in this dissertation in the
context of assessing the compliance between a CP and the medical behaviours.
Performance analysis focuses on identifying bottlenecks and measuring performance indicators
(Neira et al., 2017). Its goal is to support organizations by analyzing the performance of
their business processes and providing metrics, such as average time between events. As a
result, the information obtained can be displayed in various types of charts, like bar charts
or dotted charts.
Conformance checking inspects if the observed behaviour in the given event log conforms to the
given model (Mans et al., 2013). In other words, it evaluates how well an event log that
records the actual executions matches the model (Adriansyah et al., 2011)). It is concerned
with the detection of deviations and quantification of the discrepancies (Rojas et al., 2016).
Conformance comprises the orthogonal dimensions of fitness, precision, generalization and
structural, where the most important one for conformance is fitness, since if the fitness
17
value between the process executions and the model is low, not meaningful information
can be withdrawn from the other dimensions. It is described as the metric that checks
if the event log complies with the control flow of the model, i.e. whether the log fits the
model. If the fitness value is high, it means that the sequence of activities in an log instance
is allowed by the model. With this, and considering that healthcare organizations want
to satisfy treatment requirements but also to standardize behaviours in CPs, compliance
checking measurements ensure that HCPs’ treatment behaviours act in conformity with
the established CP conditions (Huang et al., 2014).
2.3.4 ProM Framework
The process mining techniques discussed aboved are implemented in the data mining frame-
work ProM1 (B.F. van Dongen and Aalst, 2005). ProM is an extensible open source tool that
supports an extensive range of process mining techniques and algorithms, where the event log
serves as its input. ProM was developed at the Eindhoven University of Technology and has
become the standard tool for process mining studies and the most frequently used tool for the
application of process mining techniques in healthcare (Rojas et al., 2016). When it was initially
developed, ProM only aimed to provide algorithms for the discovery of process models. Today,
it integrates various functionalities, featuring several kinds of analysis and model comparisons,
including the conformance checker that is used in this dissertation.
ProM is a modular tool (“plug-in concept”). It is possible to add new functionalities without
modifying the source code. As such, it incorporates a greatly quantity of plug-ins, and its
architecture is illustrated in Figure 2.5. ProM supports the following five types of plug-ins
(Verbeek et al., 2006):
• Mining plug-ins that take an event log and produce/construct a model.
• Conversion plug-ins to convert a model in a given format into another a different modelling
language, for instance, converting a BPMN model to a Petri Net.
• Analysis plug-ins to analyze a model, by comparing the event log with the model
• Import plug-ins to import or load an object (event log or model) from a file.
• Export plug-ins to export an object to a file.
1http://www.promtools.org
18
Figure 2.5: Overview of the ProM framework (Verbeek et al., 2006)
2.4 Evaluating Compliance to a Care Pathway
There is significant amount of work on care pathways analysis, specifically in the field of
healthcare management, where gathering insights on the actual clinical practice is receiving
increasing attention, due to the potential of the information extracted. For this purpose, given
the process model and the event log with all process executions, the objective is to output a
set of compliance metrics to measure the observed deviations. Rich literature describing how to
measure compliance and to discriminate compliant from non-compliant events has been reported.
Table 2.2 presents a brief overview of several studies that have been proposed to address CP
compliance assessment, but mainly focusing on clinical outcomes analysis.
van de Klundert et al. (2010) developed a method to define and measure deviations and
adherence, which provides insights on the pros and cons of CPs, applied to cardiac patients.
In their work, a CP is represented as a workflow process and the approach covers several CP
properties: precedence relations, partial orders, parallelism, exclusive alternatives, and nested
pathways. By describing a set of definitions, including the definition of an activity, a trajectory
(a set of ordered activities) and a feasible realization of the pathway, they present how to model
the pathway. The concept of adherence is characterized as the trajectory that a patient follows
that is a feasible realization of the pathway. The model proposed to evaluate that adherence to
the pathway using costs for the deviations measured on a numerical scale, based on the input
of clinical experts. Two combined algorithms to measure the adherence are described and the
model outputs an adherence value between 0 and 1, in which 0 refers to the perfect adherence
19
Table 2.2: Overview of CP compliance analysis studies.
Author Method Results
Quagliniet al. (2001)
Developed a system that presents aclassification of possible exceptions of aclinical guideline (a workflow process) andshowed how the sequence of tasks may bealtered.
Performed a non-compliance analysisregarding the ischemic stroke clinicalguidelines’ impact on hospital stay andcosts outcomes.
Quagliniet al. (2018)
Collected data on costs and guidelinecompliance prospectively, and examinedthe association through a multivariatestatistical model.
Discovered that treating patientsaccording to guidelines yields economicbenefits as well as a reduction in thehospital’s resources, since on average thelength of stay decreased.
Micieli et al.(2002)
Calculated a rating for each patient ofnon-compliance with the clinicalguidelines, as well as the relative risk ofdeath. Also, the influence of guidelinecompliance on disability was assessedthrough correlation tests and analysis.
Proved that there is a correlation withsurvival and treatment effectiveness withguideline compliance (in patients withfirst-ever ischaemic stroke), claiming thatEBM strongly effects the clinicaloutcomes.
Milchak et al.(2006)
Developed an adherence tool to measureindividual treatment adherence toguidelines in a case study onHypertension, combining a set of scoreson 22 criteria.
Returned a score that determines forevery patient case if each individualcriterion is met and quantified the overalladherence score.
Huang et al.(2014)
Developed a real-time approach to getinsights into CPs’ violations and supportonline compliance checking.
Provided an efficient and generalsurveillance of CPs, enabling to gatherinsights and deeper understanding intoCPs, from internal and externalperspectives.
Ainsworthand Buchan(2012)
Developed a web-based tool to representCPs as graphs and compares them todata collected on EHR to automaticallyanalyze deviations.
Returned a variance analysis in whichdifferences from the expected and actualcare are identified, by defining patterns ofdeviations.
and 1 to the complete lack of adherence. Nonetheless, the alignment (between the observed
behaviours and the model) algorithm makes it very hard to be applicable in real data since does
not account for the order of the activities (oversimplification of the reality).
Weidlich et al. (2011) presented a set of compliance measurements for the cases of a process,
not specifically in field of health. The method can be applied to any process model and event
log, modelled with the BPMN language. The proposed method relies on a set of behavioural
constraints that a process model imposes for a pair of activities, instead of trying to replay
cases accordingly to a rigorous notion of equivalence. They present an algorithm based on the
behavioural profiles of a process and a case, in which each pair of activities is in, at least, one
of the following relations: strict order relation (one activity that occurs after the other, not
necessarily adjacent), exclusive activities (activities that can not occur in a same sequence of
20
activities), interleaving relation (one activity that can occur before or after the other) or causal
relation (the occurrence of one activity enforces the occurrence of another one). The authors
also address two important issues: common noise of a process model and its impact on the
compliance and a thorough root cause analysis on non-compliance violations for a single case
and/or the entire process log. According to them, while some cases are valid executions, others
capture a certain share of the recommended behaviour, focusing on scenarios where the process
model is already given by a normative model (models that define which steps, and in which
order, have to be executed to achieve a business value).
There has been also work from the BPM community on the analysis on clinical practice
compliance to a care pathway or guideline, considering process mining techniques, in particular,
conformance checking. Rovani et al. (2015) proposed a methodology to check the conformance
of the clinical guidelines, named de jure model, against the real clinical practice in the event log,
in which the modelling language employed is a declarative language. Through a set of process
mining techniques supported by different ProM plug-ins, Declare Analyzer and the Declare
Checker, they identify the main discrepancies and provide an analysis on such deviations in the
process executions, suggesting actions to be undertaken to adapt and improve the CP. They
present a method to automatically (i) check the conformance of the de jure model, representing
the clinical guidelines, and medical behaviours. Afterwords, (ii) to repair the de jure model to be
in line with the real clinical conduct, outputting the de facto model, based on cross validation.
Their solution consists in randomly splitting the event log in two: one fragment is the training
log to repair the model and the other is the one for which the conformance is checked. If the
conformance is satisfactory, the repair model is retained and considered as the de facto model.
Alternatively, the log splitting is repeated, as well as the conformance checking and the repair,
until it achieves a satisfactory conformance value.
Adriansyah et al. (2011) present an algorithm for conformance checking given a process mod-
elled in a Petri net and the corresponding event log. The algorithm, based on the A* algorithm
(Dechter and Pearl, 1985), details a cost-based replay technique that measures deviations within
the fitness dimension. The proposed fitness value of a case (i.e. a process instance) is a high
value if the sequence of activities is feasible by the model. Such metric takes into consideration
skipping and inserting activities, in which the user must assign a cost to the individual activities
of the process, such that the severity of specific activities is accounted for in the conformance
measurement. Therefore, it provides intuitive diagnostics on the divergences according to the
process. Results on a set of experiments conducted by the authors proved that the algorithm is
suitable to measure fitness even where the completeness of the cases is not assured. Additionally,
21
it produces better results when the cost of skipping activities is higher than that of inserting
activities. In contrast to other classical fitness techniques, as the similar algorithm proposed
by Rozinat and van der Aalst (2008), the method developed by Adriansyah et al. (2011) does
not penalize the conformance for existence of either skipped or inserted activities and allows to
consider the severity of the activities. The algorithm is available from the ProM framework.
2.5 Overview
Care pathways consist in recommendations to be followed that detail structured medical
treatment processes about the appropriate care, applied to a concrete clinical setting. However,
clinical practice often deviates from these process models. Additionally, they can be out-of-
date, idealized or not even properly related to reality. Consequently, organizations pursue the
perfect balance between flexibility and control (Adriansyah et al., 2011). Thus, it is essential
to understand how the reality actually conforms with the model. Conformance checking and
compliance assessment are of utmost importance once they allow to improve the quality of the
medical treatment processes, to better design the CPs, to decrease unjustified variations and to
reduce the medical costs
The current approaches for CP’ compliance analysis mainly focus on aggregated data with
regards to the clinical outcomes, restricting the perspective of the CP and not considering the
medical behaviours, or require high-quality data to achieve meaningful results. On the other
hand, process mining is an emerging valuable discipline, gaining increasing attention in the
healthcare environment due to its potential in extracting non-trivial information. In fact, process
mining techniques offer a way to more rigorously check compliance and analyze the potential of
the information about a process (Rovani et al., 2015). To the best of my knowledge, “auditing”
on essential treatment behaviours on the execution of a pathway (at a very refined level) has
not been given enough attention, which is an important contribution with this dissertation.
This work is built on ideas from the work reviewed in this chapter. Similarly, given the
medical behaviours structured in the event log format and the CP modelled, the purpose is to
provide a set of metrics to assess and monitor the compliance of the HCPs’ performance to a
CP.
22
3Care Pathway Compliance Assessment
This chapter details the compliance assessment model proposed for evaluating and analyzing
the performance of healthcare professionals (HCP) according to a care pathway (CP). Given the
modelled CP, the event log and the severity costs set for each activity of the pathway, the model
outputs the conformity value for each patient case in association with metrics computed on
the compliance of individual activities, the order and pathway constraints, for the inspection of
the underlying non-compliant events (that supports the CP compliance analysis). Figure 3.1
illustrates the model proposed.
Section 3.1 details the event log required as input, as well as the formalization of the concepts
and non-compliance criteria used throughout the model. For this purpose, this work follows the
suggestion of Weidlich et al. (2011), as the basis to define the information’s architecture around
a CP and a patient case. The model considers an interpretation of the CP using the process
description language Business Process Management Notation (BPMN). Section 3.2 presents the
conformance checking algorithm employed to measure the conformity of clinical practice to the
CP. The proposed model adopts the work of Adriansyah et al. (2011), applied to any domain of
process models and it is available in the ProM framework. Section 3.3 presents the compliance
metrics defined to support the diagnostic analysis, based on the possible non-compliant scenarios
identified. Finally, Section 3.4 presents a summary of this chapter and its key points.
3.1 Architecture of the Model Based on Non-compliance Crite-
ria
In order to apply the data of medical behaviours on the patients careflow stored in the
information systems of an healthcare organization to process mining tools, the first stage involves
the generation of the event log table. The event log is the standard structure used to represent
the sequence of HCPs’ interventions for single patient cases.
23
Figure 3.1: Overview of the proposed model.
The event log that serves as input for the proposed model must contain three different
attributes: the clinical case identifier, the activity that occurred and the timestamp in which
the activity was performed. Each entry corresponds to a specific activity of a case and the
event log must be ordered by the timestamp. However, additional attributes can be considered,
such as the gender of the patient, the phase of the pathway, the HCP who performed the
activity, the healthcare unit in which the activity performed or any clinical outcome. With
multiple attributes, it is possible to filter the event log to study the compliance based on the
specific characteristics under analysis (e.g. filtering by the clinical cases for which mortality is
observed). Nevertheless, it must be ensured that each activity has an unique identifier, such
that the name of the activity recorded in the log match the same activity in the modelled CP
and vice versa. Moreover, the log may include different activities than those expressed in the
model.
As basis for the information’s architecture of this model, represented in Figure 3.2, the
formalization of the concepts used follows the suggestion of Weidlich et al. (2011), but applied
to medical treatment processes. Considering the authors’ preliminaries and an interpretation of
the CP using BPMN modeling language, a CP and a patient case are defined as:
(1) Definition: Care Pathway.
A care pathway is defined as CP = (A, εp, ai, ao), in which
• A is a non-empty set of activities. An activity is considered the unit of care that must be
24
Figure 3.2: Architecture of the proposed model to assess the compliance to a care pathway.
delivered to the patient, such that should be executed and recorded, e.g. interventions,
laboratory tests or prescriptions.
• εp is the set of valid executions sequences of activities.
• ai ∈ A is the initial activity and ao ∈ A is the final one, with no business semantics except
initiating and ending a process instance.
(2) Definition: Patient Case.
Following the structure of an event log, the set of patient cases is represented by C = [c1, c2,
c3, ...], where a patient case c is a list of ordered events, referring to the observed behaviours in
a patient careflow, in the form c = 〈e1, ..., en; f c〉, in which n > 0 ∈ N is the number of events
occurred, ej ∈ Ec ∀ 1 < j ≤ n and fc is the conformance value of the case. Each event has an
associated timestamp.
Under the assumption of a well-defined CP execution semantics, the existence of the gate-
ways construct in a BPMN representation is contemplated when declaring all the possible paths
of the CP. Assuming the existence of such definition, the set of feasible sequences of activities
according to the CP is presented in Definition 3. To note that an execution sequence refers to
a sequence of activities considered valid in the CP. In contrast, the sequence of events of a case
may not be completely replayed by the model.
(3) Definition: Execution Sequence.
Following the semantics of care pathway CP, the set of execution sequences for a CP is expressed
25
(a)
(b)
(c)
Figure 3.3: Missing activities of a trace: (a) missing head of a trace; (b) missing tail of a trace;(c) missing episode of a trace.
by εp = [σ1, σ2, σ3, ...], which is the set of lists, each named execution sequence, of the form σ =
〈ai, a1, ..., am, ao〉 representing a feasible sequence of activities of the CP, in which m > 0 ∈ N,
aj ∈ A ∀ 1 < j ≤ m, ai ∈ A is the initial activity and ao ∈ A is the final one.
Subsequently, to identify reasons for non-compliant executions, it has to be defined which
possible deviations can occur. With respect to the non-compliant scenarios considered, taking
inspiration on the work of Weidlich et al. (2011) and Gunther (2009), it is possible to distinguish
two main classifications: missing activities (or skipped activities) and inserted events.
• Missing activities: activities that should occur according to the CP but are not observed
in the event log. The three different types considered are illustrated in Figure 3.3.
• Inserted events: events that occurred in reality but should not be executed according to
the CP. See Figure 3.4 with the three types of these events, involving the wrong recording
of an activity order, the wrong recording of an additional event or the recording of alien
events (events not contemplated in the pathway).
3.2 Conformance Measurement
To measure the conformance value for each clinical case with the specifications in the CP,
the model uses the Replay a Log on flexible model for conformance analysis plug-in in the ProM
26
(a)
(b)
(c)
Figure 3.4: Inserted events of a trace: (a) trace with wrong order; (b) trace with additionalevent; (c) trace with alien event.
framework, developed by Adriansyah et al. (2011). The algorithm employs a replay analysis
technique that defines a fitness metric for quantifying the extent of the deviations observed,
considering the two possible causes of non-compliant scenarios – skipped activities and inserted
events.
To achieve the conformance values, the event log as well as the modelled CP must be
imported into the ProM framework. Through the help of conversion plug-ins available in ProM,
the care pathway is converted from BPMN to Petri Net, which is the required modelling language
of the algorithm. The output is a file (extracted with the help of export plug-ins), in which each
entry contains the patient case identifier and the corresponding fitness, i.e. the conformance
value.
The set of skipped activities in the case are identified as As over A. On the other hand,
all inserted events are identified as a set Ei ⊆ Ec (Ec designates the set of events of a case
c), where c is the case identifier from the finite set of cases C (c ∈ C). It is possible to take
into account different costs for skipping and inserting individual activities, depending on the
severity considered for each activity of the CP, which should be assigned by clinical experts (see
Figure 3.1). Thus, ks and ki represent the cost functions for skipping and inserting activities,
respectively. As a result, the fitness metric f is defined as: one minus the ratio between the total
cost of having inserted/skipped activities and the total cost of considering all events as inserted
activities.
27
f = 1−∑
a∈AsAs(a)× ks(a) +
∑e∈Ei
ki(α(e))∑e∈Ec
ki(α(e))(3.1)
in which α: E → A represents a function relating each event of the case to an activity of the
process.
The rationale behind this fitness metric is that the fitness values should decrease as more
activities are skipped or inserted. In the worst scenario of evaluating the conformance to a
process, all the events are inserted and that is used to normalize the fitness metric. Note that it
is possible that f becomes negative in cases where many activities are skipped and the skipping
costs are high, for instance, a loop that happens repeatedly. However, generally, f assumes
values between 0 and 1 (for a feasible sequence according to the process model).
Equation 3.1 involves knowing up front which activities are skipped or inserted. For this
reason, when identifying deviations to the CP, the focus is in discovering the inserted and skipped
activities that return the minimal cost for a valid path of the process model, such that we obtain
the maximum fitness value. Hence, identifying these activities in a case is formulated as finding
the best matching instance of the process model based on the events of a case. To achieve this,
the conformance calculation invokes the A* algorithm in order to identify the process instance
that best matches a given case. The A* algorithm was originally developed to discover the
shortest path between two nodes, the source and target nodes, in a directed graph with arc
costs (distances associated to the arcs). These “arc costs” are related to the cost functions of
skipping and inserting activities.
In detail, accordingly to the proposal of Adriansyah et al. (2011), to replay a group of events
on a process model, the instances of the process (process paths) are iteratively constructed using
shares of the sequence of events of a case. For each pair consisting of a share of the sequence of
events and a constructed instance, the events are matched to the process transitions instances
that represent the same activity. Every time a transition is fired, it refers to a matched event
in the process. Considering events in the share of events for which no match was discovered, it
means an inserted activity was found, once they occurred in reality but should happen according
to the process model. On the other hand, skipped activities are considered when process tran-
sitions instances are not associated with any event in the sequence of events, since they should
be executed according to the process model but are not observed in reality. Thus, to determine
which activities of the process instance most likely describes the deviations observed in a case,
the costs functions (ks and ki) are employed. The objective is to construct the process instance
that shows the lowest cost of deviations, such that only in cases where the cost of skipping an
28
Table 3.1: Conformance measurement between a set of cases and the pathway modelled in Figure2.3, where the cost of skipping an activity is twice than the cost of inserting one.
Case Sequence of events f value Description
1 〈A, B, C, D, L〉 1.00 Valid sequence.
2 〈A, B, C, D, E, J, H, K, L〉 0.95 Activity H was inserted.
3 〈A, B, C, D, E, F, H, G, I, L〉 0.85 Activity G was skipped and inserted.
4 〈A, B, C, E, L〉 0.80Activity D was skipped and E wasinserted.
5 〈A, B, C, X, D, E, J〉 0.71Activity L was skipped and X (alien), Eand J were inserted.
6 〈C, D, E, J, K〉 0.60 Activities A, B and L were skipped.
activity is higher than the cost of assuming an inserted event, the event is considered as inserted.
Table 3.1 presents the conformance measurement for six examples of clinical cases for the
pathway modelled in Figure 2.3, using ks=2 and ki=1. Consider the case 1, since it represents a
feasible sequence of events according to the pathway, the conformance value equals 1, in contrast
to the other cases. In case 2, activity H was considered as inserted since it is not contemplated
in that instance of the pathway. In case 3, an order violation is found where the activity G
occurred after H, therefore, it was skipped and inserted. In case 5, it is possible to observe an
inserted event, in particular an alien one – activity X. Case 6 has the lowest f value once it has
three skipped activities.
3.3 Metrics to Examine the Compliance of Patient Cases
The model comprises the definition of the compliance metrics that assist in performing the
comparison between the conformance achieved for the set of cases and the CP, which allows to
extract diagnostic information for individual patient careflow as well as for the whole population
of cases, based on the non-compliance criteria described. For this purpose, a Python script
was developed (1) to act as the link between the event log and the corresponding conformance
values. Also, (2) by implementing the concepts presented in Section 3.1, the following compliance
metrics, detailed in Table 3.2, were computed:
(a) Compliance of a patient case, i.e. the fitness value achieved, including the patient careflow
(sequence of events that was observed);
(b) Overall compliance of all patient cases (average compliance considering all cases);
29
(c) Frequency of the occurrence of an activity or set of activities, that is in how many patient
cases HCPs performed that activity or set of activities;
(d) Average compliance of the patient cases in which an activity or a set of activities has
occurred, independent of the order;
(e) Proportion of patient cases in which an activity occurred but another does not, independent
of the order, and the corresponding average compliance of those cases;
(f) Proportion of patient cases in which the order of two activities (not necessarily adjacent)
was followed as imposed in the pathway, and the corresponding average compliance of
those cases;
(g) The set of alien events, with the set of patient cases for which each inserted event occurred
and the corresponding average compliance of those cases;
(h) Proportion of patient cases and the corresponding average compliance given a particular
characteristic, e.g. a clinical outcome.
Table 3.2: Metrics computed to support the compliance analysis.
Label Metric Description Formula
a Compliance of a case c fc
b Overall compliance of all cases∑
c∈C fc∑c∈C c
cFrequency of the occurrence of a set ofactivities a
∑c∈C({a}) c∑
c∈C c
dAvg. compliance of the cases in which a set ofactivities a occurred (not considering theorder)
∑c∈C({a}) fc∑c∈C({a}) c
eProportion of the cases in which an activity aoccurred but b does not (not considering theorder) and the avg. compliance
∑c∈C({a}\{b}) c∑
c∈C c ,∑
c∈C({a}\{b}) fc∑c∈C({a}\{b}) c
fProportion of the cases in which the order oftwo activities a and b was followed and theavg. compliance
∑c∈C(a→b) c∑c∈C({a,b}) c
,∑
c∈C(a→b) fc∑c∈C(a→b) c
g Avg. compliance for each alien event
∑c∈C({i}) fc∑c∈C({i}) c
,∀i ∈ Ei
hProportion of the cases and the avg.compliance given a clinical outcome x
∑c∈Cx
c∑c∈C c ,
∑c∈Cx
fc∑c∈Cx
c
30
3.4 Summary
This chapter described the proposed model to assess the compliance of clinical practice with
a CP. Receiving as inputs the event log with, at least, the case and activity identifiers ordered by
the time of occurrence, the modelled CP in BPMN and the severity costs set for each activity of
the pathway, the model returns a set of metrics that allow to efficiently analyze the compliance
of medical behaviours to a CP.
For this purpose, the formalization of the concepts around a care pathway and a case was
defined and implemented, as well as the possible non-compliant scenarios, described as missing
activities or inserted events. The conformance measurement is leveraged with a fitness metric,
applied between any process model and the process executions. To support the compliance
analysis and the interpretation of the values obtained and of the non-compliant events, metrics
based on the frequency of occurrence of activities, compliance of individual activities, the order
and pathway constraints were computed.
With the proposed model, the purpose is not only to quantify the conformance, but to
comprehend and examine the degree of compliance of the HCPs to a CP on a population of
patient cases. It enables to account for multiple attributes (integrated in the event log) of
the population under analysis, allowing to segment the patient cases according to particular
attributes, such that different analysis can be made.
31
32
4Colorectal Cancer Care Pathway
Compliance Analysis
Colorectal cancer (CRC) is the most common malignancy of the gastrointestinal tract (Ku-
mar et al., 2015). According to the American Cancer Society1, more than fifty thousand patients
die of this disease each year in the United States, which makes CRC the third most deadly cancer
in the United States. The incidence of CRC is slightly higher in men than in women and it is
more frequent in age group of 60-70 years old, with less than 20% of the cases occurring before
50 years old. It is prevalent in the United States, Canada, Australia, New Zealand, Denmark,
Swedish amongst other developed countries. On the other hand, India, South America and
Africa reveal the lowest levels of incidence.
The adoption of screening programs is progressively decreasing the incidence of CRC (Kumar
et al., 2015). An early detection and improvements in medical and surgical care are responsible
for the observed reduction in mortality in the past few years. The identification of risk factors
(e.g. aging, hereditary factors, dietary factors, alcohol intake, cigarette smoking, etc.) for the
development of colorectal cancer is fundamental in the follow-up of susceptible populations to
the disease.
This chapter describes the evaluation of compliance between the CRC care pathway (CP)
and the medical behaviours of healthcare professionals (HCP) from Hospital da Luz Lisboa,
using the proposed model. Section 4.1 presents a brief statistical characterization of the CRC
dataset and describes the dataset preparation to achieve the required inputs for the proposed
model. Section 4.2 details on the overall compliance results obtained. Section 4.3 provides
further insights to interpret the CRC compliance values achieved and presents the associations
between the compliance results and the clinical outcomes recorded by the hospital. Finally,
Section 4.4 summarizes the chapter and the obtained results.
1https://www.cancer.org/
33
Table 4.1: Statistical characterization of the colorectal cancer dataset.
Number of entries 345
Number of variables 229
Number of fields values 79 005
Non-empty fields values (27.2%) 21 487
Empty fields values (72.8%) 57 518
Colon cancer patients 246
Rectum cancer patients 99
Colon and rectum patients 18
Male patients 208
Female patients 137
Age (minimum/average/maximum) 38/69/94 years
Average time between entering in the institution and the first treatment 20.2 days
Average time between Post-operative and discharge 7.1 days
Number of death patients 6.4%
Cardiovascular comorbidity 27.8%
Respiratory comorbidity 11.3%
Endocrine comorbidity 1.5%
Smoking comorbidity 2.9%
4.1 Dataset Preparation
The dataset assessed consists of data on colorectal cancer patients from Hospital da Luz
Lisboa entering in the hospital in the year of 2017. In the dataset, each instance/entry represents
a clinical case and presents information considering the three phases of the pathway: diagnosis,
staging and treatment. A brief statistical profile of the dataset is given in Table 4.1.
For multiple reasons, as expressed by van de Klundert et al. (2010), the data required to
measure the conformity between the actual treatment behaviours and the guidelines or care
pathways is often incomplete or incorrect, which makes compliance checking a difficult task. In
fact, only about 27% of the CRC dataset is complete and 117 variables (in a total of 229) are
empty, not considered in this CP analysis. Considering the CRC dataset, 60.3% of the patients
are male (39.7% female patients), thus reflecting a higher incidence of men than in women. The
average age for the patients population is 69 years old, within the range of ages in which CRC
is frequently observed.
The dataset reflects information about 18 key performance indicators (KPI) that the hospital
is collecting on the CRC CP. Therefore, the dataset represents the activities of the pathway that
the hospital has interest in monitoring. However, the data under analysis does not reflects the
data recorded in the EHR system, since it does not include all medical behaviours. For each
34
patient case, I considered only the activities of the pathway matching available events recorded
in the dataset. Based on these activities, I modelled two pathways with the BPMN modelling
language, one for the colon and another for the rectum. Accordingly, the dataset was split into
two subsets of patient cases, for the colon (246 cases) and rectum cancer patients (99 cases).
Since the CRC CP of Hospital da Luz Lisboa is confidential, the set of activities that form
the CP and their sequence are not revealed throughout this dissertation. Due to the nature of
the dataset, for the whole CRC pathway, I considered a total of 15 activities (based on available
events), each labeled from A to O. Each activity of the pathway was assigned to a clinical
category, including: registries, exams, laboratory tests, surgery, treatment and time constraints.
With the set of activities distinctly modelled, with regards to the event log preparation, a Python
script was developed to automatically generate the two event logs from the data collected. The
costs assigned to skipped activities were twice the costs for inserted activities considering than
skipping a mandatory activity should be more penalized than inserting one (wrong order or alien
event). These costs were chosen to carry an objective analysis, not contemplating any severity
level for different activities.
To note that, due to the nature of the dataset, during the process of the event log prepara-
tion, the following assumptions were made:
1. If a field value is empty, I assume the activity has not occurred.
2. The process modeling and event log generation are limited to the non-empty variables in
the dataset. Some patients may have performed CP activities outside the scope of this
hospital.
3. The patients containing the colon and rectum cancers were considered as two clinical cases,
one for each pathway.
4. Particular results of exams, lab tests or specific diagnostics were not addressed.
5. Due to the nature of the dataset, date and time is not available for all events studied.
Therefore, it becomes impossible to assess the order of all the observed events. For this
reason, the order of the events without a timestamp was assumed to be the order stated
on the CP. On the other hand, events with a timestamp were ordered normally.
6. Order was only assessed for consecutive activities with a timestamp. One could have
also assessed order of all timestamped activities (consecutive and non-consecutive) by
indexing the activities without timestamps to the timestamped ones, according to the
protocol. However, this would have doubled inaccurately the inserted and skipped, without
supporting evidence.
35
Table 4.2: Proportion of patient cases arranged by the number of events occurred.
Patient Cases
Events Colon Rectum Total
2 6.9% 7.1% 7.0%
3 6.1% 4.0% 5.5%
4 8.5% 20.2% 11.9%
5 20.0% 25.3% 22.9%
6 23.6% 22.2% 23.2%
7 17.5% 11.1% 15.7%
8 12.2% 8.1% 11.0%
9 3.3% 2.0% 2.9%
Figure 4.1: Distribution of patient cases by the number of events occurred.
4.2 Overall Compliance Results
Table 4.2 presents the patient cases grouped by the number of events executed for each
patient careflow. Results range from 2 to 9 events, for both subsets, where the number of events
per case for which the relative frequency is higher is 5 and 6 events, observed in 22.9% and
23.3% of the clinical cases, respectively. One can already conclude that there are multiple cases
that did not completely followed the CRC pathway.
Table 4.3 details the alignment results, considering the number of events observed and the
skipped and inserted activities, for each subset of patient cases. Considering the 99 rectum
patient cases, a total of a 522 events were observed. On the other hand, the 246 patients with
colon cancer revealed 1394 events. As expected, due to the nature of the dataset available for
this study, the number of skipped activities is considerably higher than that of inserted events.
With respect to the inserted events and considering the assumptions made, they concern only
the timestamped activities and alien events. The average events per case for the colon subset was
36
Table 4.3: Overall alignment results.
Events/case
SubsetPatient
Cases
Total
EventsAvg. Max. Min. Std. Dev.
Skipped
Activities
Inserted
Events
Colon 246 1394 5.67 (56.7%) 9 (90.0%) 2 (20.0%) 1.75 1210 141
Rectum 99 522 5.27 (52.7%) 9 (90,0%) 2 (20.0%) 1.63 507 34
Total 345 1916 5.56 9 2 - 1717 175
Table 4.4: Overall compliance results.
SubsetPatient
CasesAvg. Compliance Max. Min. Std. Dev. 1st Quartile Median 3rd Quartile
Colon 246 0.585 0.931 0.273 0.142 0.500 0.577 0.667
Rectum 99 0.573 0.931 0.273 0.139 0.500 0.577 0.673
Total 345 0.582 0.931 0.273 0.141 0.500 0.577 0.667
5.67, which represents 56.7% of the number of events for the shortest valid sequence according the
colon pathway (10 activities), showing that on average only half of the CP is completed. Similar
results were obtained for the rectum subset, with an average of 5.27 events per patient case,
which corresponds to 52.7% of the number of events for the shortest valid sequence according
the rectum pathway (10 activities).
Table 4.4 presents the overall scores obtained for the compliance of the HCPs in the treat-
ment of CRC. The values range from 0, in case of complete lack of conformance (skipping or
inserting all activities), and 1, for feasible sequences of events according to the pathway. The
overall compliance score observed was 58.5% and 57.3% for the colon and rectum patient cases,
respectively. Both subsets recorded a maximum compliance of 93.1%, in cases where only one
activity is missing, and a minimum of 27.3%, for cases in which only two activities were observed,
as illustrated in the distributions in Figure 4.2. Therefore, no patient case has data evidencing
full compliance with the CP.
4.3 Compliance Evaluation and Interpretation
4.3.1 Patients Distribution by Compliance Levels
Figure 4.3 illustrates the distribution of the patients cases according to their group of com-
pliance value, defined as equally divided groups within the observed range of compliance values
37
Figure 4.2: Distribution of the compliance values for each patient case.
(max. and min. observed) of the total patient cases.
• Low compliance group: [min, min+max−min3 ] = [0.273, 0.492].
• Moderate compliance group: [min+max−min3 , min+2* max−min
3 ] = [0.493, 0.712].
• High compliance group: [min+2* max−min3 , max ] = [0.713, 0.931].
One can conclude that the larger proportion of patient cases have moderate compliance values,
but there is still much room for improvement in increasing the compliance to the CRC pathway,
since only 15.9% of the patient cases fall into the high compliance group.
4.3.2 Compliance for Each Activity of the Pathway
When inspecting the compliance to a care pathway, several cases where a mandatory activity
of the process is often skipped possibly indicate that the CP must be adjusted to include such
skipping. Otherwise, the hospital should impose new mechanisms to enforce the best clinical
practice, expressed in the CP. Table 4.5 reports on the relative frequency of occurrence for each
activity of the pathway, for both subsets of patient cases. Some activities are only included in
one of the pathways, marked in the table with “-”.
The activities where higher frequencies were observed are in the categories of Registries,
followed by Appointment, Time and Surgery. On the other hand, lower frequency categories
involved Treatment, Exams and Lab Tests. Activities under the category of Treatment are in-
volved in alternative paths of the pathway, which explains the lower compliance values obtained.
Note that, it is not unusual for some patients to move to different healthcare units to perform
38
Figure 4.3: Distribution of all patient cases by the group of compliance values: low, moderateor high.
CP activities, which is not being considered in this analysis. Activity “G” is a mandatory activ-
ity in the pathway and a relevant KPI monitored by the hospital that refers to the solicitation
of the Carcinoembryonic Antigen (CEA), a laboratory test that must be requested before the
treatment to define the basal value to monitor recurrence signs, in the post-treatment period.
Although it is not confirmed, there has been recorded an association between higher levels of
CEA and a worst prognosis (Destri et al., 2015). In the dataset, the activity is observed only in
47% colon patient cases and 31% in the rectum cases, which may indicate a lack of awareness
of the pathway. Another example is that only 70% of the cases recorded the surgery activity,
which is mandatory according to the colon cancer pathway.
4.3.3 Compliance by the Phase of the Pathway
The medical behaviours and the activities considered in the CP can also be arranged accord-
ing to the three phases of the colorectal cancer pathway: diagnostic, staging and treatment. The
compliance observed for each phase is presented in Table 4.6. Results showed that the critical
phase is staging, for both patient subsets, since it has the lowest average frequency levels: 29.2%
and 27.8% for colon and rectum patient cases, respectively. A possible reason for these low
values lies on the low percentages of occurrence for the activities “E” (rectum) and “F” (colon),
both contained in the staging phase and referring to the set of exams that must be requested to
define the clinical stage of the patient before initiating the treatment (as observed in Table 4.5).
39
Table 4.5: Frequency of occurrence for each activity of the CRC care pathway
Colon Patient Cases Rectum Patient Cases
Activity Category Frequency Frequency
A Registry 100.0% 100.0%
B Registry 100.0% 100.0%
C Exams 21.1% 6.1%
D Time 13.8% 10.1%
E Exams - 13.1%
F Exams 18.3% -
G Lab. Tests 47.2% 31.3%
H Registry 37.4% 56.6%
I Time 68.3% 68.7%
J Appointment 87.4% 86.9%
K Surgery 70.3% 46.5%
L Treatment - 1.0%
M Treatment - 4.0%
N Treatment - 1.0%
O Treatment 1.2% 2.0%
Table 4.6: Average relative frequency of occurrence for completing any of the activities of eachphase of the colorectal cancer pathway.
Diagnostic Staging Treatment
Subset Avg. Freq. Avg. Freq. Avg. Freq.
Colon 73.7% 29.2% 56.8%
Rectum 68.7% 27.8% 30.0%
4.3.4 Compliance of the Inserted Events of the Pathway
With regards to inserted activities, in particular to alien events, i.e. events that occurred in
the treatment of these CRC patients but are not contemplated in the pathway, their importance
is essential in assessing the performance and resources management but also in redesigning the
CP. Considering an alien event with high frequency of occurrence with low average compliance
on those cases, it indicates a possible systematic mistake that must be addressed. On the other
hand, if the alien event has high frequency with high compliance values, it probably indicates
that it is an activity that should be contemplated in the CP. Due to the nature of the information
in the dataset and considering the assumptions made, the alien events were found exclusively in
the colon subset because only the events that can be observed in the rectum pathway but not
40
Table 4.7: Compliance results on the alien events observed for the colon cancer patients.
Alien Event Frequency Avg. Compliance Compliance Group
L 0.4% 0.679 Moderate
M 0.8% 0.518 Moderate
N 0.4% 0.643 Moderate
Total 1.6% 0.589 Moderate
in the colon one were considered to be alien events. Table 4.7 presents the compliance results
on the alien events observed.
Although we have an extremely reduced number of alien events, some conclusions can be
drawn from the content of the Table 4.7. Firstly, the total average compliance obtained for these
cases with alien events is above the median which sets the ground for further analysis, and also no
patient case belong to the low group of compliance values. With the help of the metrics defined,
the details on the careflow of these patient cases were discovered. The three alien events observed
correspond to treatment alternatives of the CP. According to the pathway, the treatment plan
of the colon and rectum patients must always be discussed in the multidisciplinary appointment
(MDA). Additional inspection on the sequence of events occurred revealed that there was not a
MDA in the treatment with the “L” activity. Considering the other cases, although the MDA
occurred, the treatment of colon cancer with “M” or “N” bypassed the decision process of a
MDA, which indicates that the plan was not formally discussed. One of these patient cases
is part of the group of patients that contains both cancers, therefore, the alien event can be
justified by the rectum pathway. Although the number of alien events is reduced, more than
one violation was found, which, can indicate a misunderstanding between the pathways of the
colon and rectum. Either way, a closer monitoring should be given to the way the treatment for
colon cancer is being performed.
When analyzing the frequency of occurrence for which the surgery occurred after the treat-
ment plan discussed in the MDA, as described in the pathway, it is possible to conclude that:
• For the colon cancer patient cases, the MDA and surgery occurred in 64.3%, where only
13.8% have performed those activities in the correct order.
• For the rectum cancer patient cases, both activities were observed in 41.4%, where only
22.0% have performed them in the correct order.
41
Table 4.8: Average compliance on the post-operative days until hospital discharge.
Post-operative days until Hospital Discharge
Subset Avg. Days Avg. Compliance Max. Min. 1st Quartile Median 3rd Quartile
Colon 6.1 0.631 0.931 0.375 0.577 0.600 0.750
Rectum 9.6 0.628 0.931 0.375 0.485 0.588 0.692
Total 6.7 0.631 0.931 0.375 0.577 0.600 0.750
Figure 4.4: Average post-operative days until hospital discharge per compliance group for thecolon subset of patients cases.
4.3.5 Clinical Outcomes Analysis
With respect to clinical outcomes, the dataset provided includes information on post-
operative days until hospital discharge and mortality. The first refers to the number of days a
patient was hospitalized after surgery. From the dataset of 345 patient cases, 219 were submitted
to surgery, although only 94 have records of post-operative days until discharge. The 18 patients
containing the colon and rectum cancers were excluded from this analysis since they have two
compliance values, due to the assumption made. Therefore, from the 94 patient cases, only n=88
were considered for this outcome. Table 4.8 presents the summary for this outcome and Figures
4.4 and 4.5 are the graphical representation of the average post-operative hospitalization days
until discharge and the compliance per group. Both figures show that the group with higher
compliance levels recorded, on average, has reduced days in the post-operative period, against
the other groups.
To discover the actual association between the variables, statistical tests were conducted.
The post-operative days until hospital discharge per case, for both subsets and the corresponding
compliance group (low, moderate, high) were submitted to the Shapiro test to find if the variables
follow a normal distribution, where p-value=0.018<α=0.05 (CI=95%), the significance range
42
Figure 4.5: Average post-operative days until hospital discharge per compliance group for therectum subset of patients cases.
Table 4.9: Kruskal-Wallis results on the post-operative days until hospital discharge per groupof compliance.
Compliance Group Kruskal-Wallis Ranks
Frequency Mean Rank
Low 20.5% 53.9
Moderate 51.1% 47.4
High 28.4% 32.6
Total 100.0% -
considered (Panagiotakos, 2008). Proved otherwise, the non-parametric test Kruskal-Wallis
was applied, where p-value=0.013<α=0.05 (CI=95%). By rejecting the null hypothesis of the
equality of the distribution between the three groups under analysis, it is possible to conclude
that there is an association between the compliance level and the number of post-operative days,
for at least one of the groups. By analyzing the mean ranks in Table 4.9, there is statistical
significance that the number of post-operative days is lower for patient cases that showed higher
levels of compliance.
The second outcome analyzed was the mortality and the compliance results are presented
in Table 4.10, where it can be observed that all statistical values are lower than those for the
whole population of cases. From the sample of patient cases, the patients with both cancers
were excluded, having a total of n=309 patient cases, for which there are records of 18 deaths.
Note that are only being considered the deaths recorded by the hospital. However, it is possible
that a patient died out of the scope of the hospital. To understand if there is an association
between the group of compliance and the mortality, the Chi-square test was conducted. This
43
Table 4.10: Average compliance on the mortality.
Mortality
SubsetAvg.
ComplianceMax. Min. 1st Quartile Median 3rd Quartile
Colon 0.499 0.692 0.273 0.436 0.500 0.588
Rectum 0.539 0.692 0.273 0.500 0.500 0.596
Total 0.514 0.692 0.273 0.480 0.500 0.594
test analyzes the independence of two categorical variables, proposing as the null hypothesis that
the variables are independent. Results show that p-value=0.141>α=0.05 (CI=95%), rejecting
that there is an association between mortality and the group of compliance, i.e. the variables
under analysis are independent.
The compliance analysis is essential to evaluate and understand the conformance of the
HCPs to the pathways. How the medical behaviours are reflected in the patients’ clinical out-
comes must also be addressed. It is fundamental to confirm if, by enforcing the adoption of
the pathway or by improving and adjusting the pathway to the reality of clinical practice, the
hospital achieves better clinical results. Actually, some negative discrepancies can be avoided if
HCPs are aware of the corresponding outcomes.
4.4 Summary
This chapter presented the assessment of the compliance between the CRC CP and the
observed treatment behaviour from Hospital da Luz Lisboa. The dataset was recognized to be
highly uncompleted, 73% is empty and manifesting information only around the KPIs the hos-
pital is collecting, increasing the complexity on the task of a complete CP compliance checking.
Due to the nature of the data, the dataset on the hospital’s CRC CP was prepared resulting
in two the event logs, one for the colon cancer and the other for the rectum cancer, and two
pathways modelled with the activities considered. The results obtained for the average events
per case were 57% and 52%, for the colon and rectum pathway, showing that on average only
about half of the pathway was completed. The overall compliance values obtained for the colon
and rectum subset of patient cases was 58.5% and 57.3%, respectively, having a total of 1916
events, 1717 missing activities and 175 inserted events.
Reports on the CRC pathway compliance allowed to conclude that 15.9% of the clinical cases
fall into the group of higher compliance values (considering the interval of observed values). Also,
44
in the analysis conducted, it was possible to address the frequency of occurrence and the average
compliance for individual activities. Special attention has been given to critical activities, with
lowest levels of occurrence, and results exhibited that the staging phase of the pathway should
be carefully monitored. Additionally, several violations were found, where one could conclude
that the treatment plan is frequently not discussed in the multidisciplinary appointment, as
described in the CRC pathway.
As an attempt to better understand the impact of compliance to the CRC CP in the patients’
life, associations between the compliance levels and the clinical outcomes recorded by the hospital
were made. Considering post-operative days until hospital discharge, results showed there is
statistical significance that the number of post-operative days is lower for the patient cases that
showed higher levels of compliance. For mortality, results showed that there is no association
between mortality and the group of compliance.
45
46
5Conclusions and Future Work
This dissertation presented a model to assess the compliance between the clinical practice
and a care pathway (CP), applied to a real-world dataset on colorectal cancer (CRC) from
Hospital da Luz Lisboa. This chapter summarizes the main conclusions and contributions of
this work and outlines possible directions for future work.
5.1 Conclusions
This study was motivated by the current gap that exists between the clinical practice and the
recommendations in a care pathway. Care pathways and clinical practice guidelines have been
proven to be beneficial structured processes about the appropriate healthcare and, therefore,
are widely used around healthcare organizations. However, they are often idealized processes
that are not always followed by healthcare professionals (HCP), such that implementing a care
pathway and its adoption represent a critical challenge to the healthcare organizations. In
fact, scientific literature revealed that simply implementing a CP itself is not enough. Much
attention must be paid to the design and improvement of these pathways. The need to monitor
all the activities executed and to identify the main non-compliant events is constantly increasing
as an attempt to deliver treatment excellence, reduce unjustified deviations and leverage cost-
efficiency in the quality of care services. Therefore, this model emerged for the assessment of
the compliance between the clinical practice and the CRC care pathway from Hospital da Luz
Lisboa, supporting process improvement.
In scope with the objectives set for this work, the conducted research was mainly focused
in understanding the healthcare environment and in identifying which algorithms and methods
are currently being employed to measure compliance, having the patient careflow and the CP
modelled a priori. Based on relevant literature, process mining techniques were selected due to
their potential in extracting non-trivial information out of processes executions. In particular,
a conformance checking algorithm was applied, within the fitness dimension, that replays a
47
sequence of events in the CRC care pathway. ProM was the framework employed, which has
become the standard tool for process mining studies and is the most common used one in
healthcare.
Given the modelled CP using BPMN (along with its executions sequences), the event log
and the activities’ costs, the model outputs the conformity value (a fitness metric) for each
patient case in line with computed metrics based on non-compliant scenarios defined to support
the compliance analysis. There are some difficulties in the conformance checking for medical
treatment processes mainly introduced by the richer diversity of medical behaviours over other
business processes. However, the model proved to be a robust approach once it allowed to
deliver valuable information of a dataset that was recognized as highly incomplete. Results
showed that no patient case has data evidencing full compliance with the CP, i.e. there is no
sequence of events that is completely feasible by the process. It was observed that there is still
much space for improvement since on average only about half of the pathway was completed.
In detail, the activities of the staging phase recorded the lowest occurrence levels, possibly
related with incorrect solicitation staging exams, reason why should be carefully monitored by
the hospital. Additionally, the number of skipping activities was significantly higher than the
number of inserted events. Considering the latter ones, results showed that the treatment plan
frequently bypassed the multidisciplinary appointment, which represents a significant violation
of the pathway. It is important to refer that the scenarios where patients moved to other hospital
units was not addressed in the analysis.
Concerning the clinical outcomes analysis for the CRC population, as an attempt to better
understand the impact of compliance in the patients’ life, possible associations between the vari-
ables under analysis were discovered. Considering post-operative days until discharge, results
showed that there is an association between the number of post-operative days and the compli-
ance level for at least one of the groups. As suggested in the results, patient cases that revealed
high levels of compliance were hospitalized during less days, on average. For the mortality, no
association was found between mortality and the groups of compliance. It is important to un-
derstand that one can not mislead on the subjectivities involved in the treatment interventions,
as the complexity of the patients cases, which can interfere with the results obtained.
5.2 Future Work
I ascertain that data quality is essential for the success of an optimal alignment search and
the main limitation of this study is related with the availability of event logs that record the
48
executions of patients cases. In this work, the patient data considered manifested information
around the key performance indicators (KPI) the hospital is collecting on CRC, rather than
the actual medical interventions along or around the CP. In addition, not having the date of
occurrence for every event executed does not allow to consider the correct order of events, which
reflects a limitation of the data. Judging by the high percentage of empty field values in the
CRC dataset, considering that an activity has not occurred if a field value is empty influences
the results because, in reality, the activity could have been performed but not recorded. In
further studies, a larger population of patient cases containing all clinical behaviours and their
timestamps (complete event logs) should be considered, in order to contemplate all activities of
the pathway.
With respect to the compliance analysis, even though valuable results were achieved, there
are multiple possibilities for future work. As conducted by van de Klundert et al. (2010), depend-
ing on the characteristics of individual activities in the CP, severity should also be addressed.
As future improvements, based on the input of clinical experts, defining different “costs” for
skipping and inserting each activity explores the real-life scenario in measuring how the clinical
executions comply with the model. Further improvements rely on a method to address and
classify the complexity of patients cases, for instance as in the proposal of Schaink et al. (2012),
such that it can be contemplated in the compliance measurement, as an attempt to have a closer
approach for the actual healthcare environment.
The proposed model leverages process mining techniques to measure the conformance be-
tween a CP and an event log, based on a replay technique, but other types of approaches have
been published. For instance, the proposal of Weidlich et al. (2011) consists in leveraging causal
behavioural profiles instead of an exact replay of the cases according to the notion of equivalence.
The authors argue that such profiles are a behavioural abstraction, which is more relaxed than
trace equivalence between a process execution and a process model.
Another idea worth exploring would be to provide continuous monitoring and real-time
feedback on the conformance of the patient careflow according to a CP regulations. Huang
et al. (2014), adopted experimental techniques that provide online parallel compliance checking
on the treatment constraints imposed by the CP specifications. In their approach, treatment
constraints are modeled using Semantic Web Rule Language (SWRL) rules to characterize the
clinical practice and their specific temporal formulation. Such SWRL rules are integrated with
the CP workflow such that compliance checking is applied during the pathway execution. On the
other hand, process mining techniques seldom provide measurements on temporal information
that are often contemplated in the CP. As an added topic for further analysis, compliance
49
checking on the time intervals enables to perceive the temporal impact in CPs that differ in
time spreadings (Huang et al., 2012).
In terms of the practical application of this dissertation, although the model was proposed
in line with the information of the dataset provided on the CRC, to the best of my knowledge,
the model is suitable to assess the compliance between any care pathway and corresponding
medical executions. This model provides a proper monitoring of the clinical practice that can
work as an auditing model for process improvement and quality care assurance. Deviations
for all care pathways within an hospital can be measured and effortlessly gather insights on
the HCPs performance (according to the CP). In fact, in conducting such analysis, the value
is not in the compliance measurement itself, but in the improvements in the care provided.
Also, to scale to the whole set of care pathways of an hospital, as well as to apply the proposed
methodology in different scenarios within multiple departments or units, is considered a relevant
application. In line with this, this explicit applicability sets the foundation for the development
of a platform that receives the event log and allows to model the care pathway, returning the
conformity measurements and providing patient data dashboards to aid in the visualization of
the compliance information and clinical outcomes.
50
Bibliography
Adriansyah, A., van Dongen, B. F., and van der Aalst, W. M. P. (2011). Conformance Check-
ing Using Cost-Based Fitness Analysis. In Proceedings of the 2011 IEEE 15th International
Enterprise Distributed Object Computing Conference, EDOC ’11, pages 55–64, Washington,
DC, USA. IEEE Computer Society.
Ainsworth, J. and Buchan, I. (2012). COCPIT: A tool for Integrated Care Pathway Variance
Analysis. Studies in Health Technology and Informatics, 180:995–999.
Basse, L., Hjort Jakobsen, D., Billesbølle, P., Werner, M., and Kehlet, H. (2000). A Clinical
Pathway to Accelerate Recovery After Colonic Resection.
B.F. van Dongen, A.K.A. de Medeiros, H. V. a. W. and Aalst, W. V. D. (2005). The ProM
framework: A new era in process mining tool support. Applications and Theory of Petri Nets
2005, 3536/2005(i):444–454.
Cornu, J.-N., Cussenot, O., Haab, F., and Lukacs, B. (2010). A Widespread Population Study
of Actual Medical Management of Lower Urinary Tract Symptoms Related to Benign Pro-
static Hyperplasia Across Europe and Beyond Official Clinical Guidelines. European Urology,
58(3):450–456.
Dechter, R. and Pearl, J. (1985). Generalized best-first search strategies and the optimality af
A*. Journal of the ACM, 32(3):505–536.
Destri, G. L., Rubino, A. S., Latino, R., Giannone, F., Lanteri, R., Scilletta, B., and Di Cataldo,
A. (2015). Preoperative carcinoembryonic antigen and prognosis of colorectal cancer. An
independent prognostic factor still reliable. International Surgery, 100(4):617–625.
Dongen, V. and Frans, B. (2007). Process mining and verification. Dissertation abstracts inter-
national, 68.
Every, N. R., Hochman, J., Becker, R., Kopecky, S., and Cannon, C. P. (2000). Critical Path-
ways. Circulation, 101(4):461 LP – 465.
Gorp, P. V. and Dijkman, R. (2013). A visual token-based formalization of BPMN 2.0 based on
in-place transformations. Information and Software Technology, 55(2):365–394.
51
Gunther, C. W. (2009). Process Mining in Flexible Environments.
Hajjaj, F. M., Salek, M. S., Basra, M. K. A., and Finlay, A. Y. (2010). Nonclinical influences,
beyond diagnosis and severity, on clinical decision making in dermatology: Understanding the
gap between guidelines and practice. British Journal of Dermatology, 163(4):789–799.
Hornix, P. T. (2007). Performance analysis of business processes through process mining. Mas-
ter’s Thesis, Eindhoven University of Technology, (January):1–108.
Huang, Z., Bao, Y., Dong, W., Lu, X., and Duan, H. (2014). Online treatment compliance
checking for clinical pathways. Journal of medical systems, 38(10):123.
Huang, Z., Dong, W., Ji, L., and Duan, H. (2016). Predictive monitoring of clinical pathways.
Expert Systems with Applications, 56:227–241.
Huang, Z., Lu, X., and Duan, H. (2012). On mining clinical pathway patterns from medical
behaviors. Artificial Intelligence in Medicine, 56(1):35–50.
Huang, Z., Lu, X., Duan, H., and Fan, W. (2013). Summarizing clinical pathways from event
logs. Journal of Biomedical Informatics, 46(1):111–127.
Kumar, V., Abbas, A. K., and Aster, J. C. (2015). Robbins Patologia Basica. Elsevier.
Lenz, R. and Reichert, M. (2007). IT support for healthcare processes - premises, challenges,
perspectives. Data and Knowledge Engineering, 61(1):39–58.
Macario, A., Horne, M., Goodman, S., Vitez, T., Dexter, F., Heinen, R., and Brown, B. (1998).
The effect of a perioperative clinical pathway for knee replacement surgery on hospital costs.
Anesth Analg, 86(5):978–984.
Mans, R. S., Vanwersch, R. J. B., and van der Aalst, W. (2013). Process Mining in Healthcare:
Opportunities Beyond the Ordinary. Bpmcenter.Org, pages 1–52.
Marrie, T. J., CY, L., SL, W., and et al (2000). A controlled trial of a critical pathway for
treatment of community-acquired pneumonia. JAMA, 283(6):749–755.
Micieli, G., Cavallini, A., and Quaglini, S. (2002). Guideline compliance improves stroke out-
come: A preliminary study in 4 districts in the Italian region of Lombardia. Stroke, 33(5):1341–
1347.
Milchak, J. L., Carter, B. L., Ardery, G., Black, H. R., Bakris, G. L., Jones, D. W., and Kreiter,
C. D. (2006). Development of explicit criteria to measure adherence to hypertension guidelines.
Journal Of Human Hypertension, 20:426.
52
Neira, R. A. Q., De Vries, G. J., Caffarel, J., and Stretton, E. (2017). Extraction of data from
a hospital information system to perform process mining. Studies in Health Technology and
Informatics, 245:554–558.
Panagiotakos, D. B. (2008). Value of p-value in biomedical research. The open cardiovascular
medicine journal, 2:97–99.
Panella, M., Marchisio, S., and Di Stanislao, F. (2003). Reducing clinical variations with clinical
pathways: do pathways work? International Journal for Quality in Health Care, 15(6):509–
521.
Quaglini, S., Cavallini, A., Gerzeli, S., and Micieli, G. (2018). Economic benefit from clinical
practice guideline compliance in stroke patient management. Health Policy, 69(3):305–315.
Quaglini, S., Stefanelli, M., Lanzola, G., Caporusso, V., and Panzarasa, S. (2001). Flexible
guideline-based patient careflow systems. Artificial Intelligence in Medicine, 22(1):65–80.
Rebuge, A. and Ferreira, D. R. (2012). Business process analysis in healthcare environments: A
methodology based on process mining. Information Systems, 37(2):99–116.
Rojas, E., Munoz-Gama, J., Sepulveda, M., and Capurro, D. (2016). Process mining in health-
care: A literature review. Journal of Biomedical Informatics, 61:224–236.
Rovani, M., Maggi, F. M., De Leoni, M., and Van Der Aalst, W. M. (2015). Declarative process
mining in healthcare. Expert Systems with Applications, 42(23):9236–9251.
Rozinat, A. and van der Aalst, W. M. P. (2008). Conformance checking of processes based on
monitoring real behavior. Information Systems, 33(1):64–95.
Schaink, A., Kuluski, K., Lyons, R., Fortin, M., Jadad, A., Upshur, R., and Wodchis, W.
(2012). A scoping review and thematic classification of patient complexity: offering a unifying
framework. Journal of Comorbidity, pages 1–9.
van de Klundert, J., Gorissen, P., and Zeemering, S. (2010). Measuring clinical pathway adher-
ence. Journal of Biomedical Informatics, 43(6):861–872.
van den Bos, J., Rustagi, K., Gray, T., Halford, M., Ziemkiewicz, E., and Shreve, J. (2011).
The $17.1 billion problem: The annual cost of measurable medical errors. Health Affairs,
30(4):596–603.
53
van der Aalst, W. M., Reijers, H. A., Weijters, A. J., van Dongen, B. F., Alves de Medeiros, A. K.,
Song, M., and Verbeek, H. M. (2007). Business process mining: An industrial application.
Information Systems, 32(5):713–732.
van der Aalst, W. M. P., ter Hofstede, A. H. M., and Weske, M. (2003). Business Process Man-
agement: A Survey. Business Process Management. BPM 2003. Lecture Notes in Computer
Science, 2678:1–12.
Verbeek, H. M., Van Dongen, B. F., Mendling, J., and Van Der Aalst, W. M. (2006). Interop-
erability in the ProM framework. CEUR Workshop Proceedings, 200.
Weidlich, M., Polyvyanyy, A., Desai, N., Mendling, J., and Weske, M. (2011). Process compliance
analysis based on behavioural profiles. Information Systems, 36(7):1009–1025.
Weiland, D. E. (1997). Why use clinical pathways rather than practice guidelines? The American
Journal of Surgery, 174(6):592–595.
Yan, H., Gorp, P. V., Kaymak, U., Lu, X., Ji, L., Chiau, C. C., Korsten, H. H. M., and Duan, H.
(2018). Aligning Event Logs to Task-Time Matrix Clinical Pathways in BPMN for Variance
Analysis. IEEE Journal of Biomedical and Health Informatics, 22(2):311–317.
Yan, H., Van Gorp, P. M., Kaymak, U., Ji, L., Lu, X., Chiau, C. C., Korsten, H. H., and
Duan, H. (2017). Aligning Event Logs to Task: Time Matrix Clinical Pathways in BPMN for
Variance Analysis. International Conference on Biomedical and Health Informatics (BHI),
2194(c):1–7.
Yang, W. and Su, Q. (2014). Process mining for clinical pathway: Literature review and future
directions. In 2014 11th International Conference on Service Systems and Service Management
(ICSSSM), pages 1–5.
54
AProM – Conformance Checking
Figures A.1 and A.2 represent the replay results of the conformance between the CRC
event log and the CRC CP, on the ProM framework. These two ProM displays show the events
performed by the HCPs for every patient case, grouped by identical sequences of events, including
the skipped activities and inserted events.
Figure A.1: Conformance checking between the rectum patients careflow and the pathway.
55
Figure A.2: Conformance checking between the colon patients careflow and the pathway.
56
Recommended