A Process Mining Framework to Analyze Variability in Human Behavior

Gemma Di Federico

Research output: Book/ReportPh.D. thesis

14 Downloads (Pure)

Abstract

Over the years, process mining has become increasingly recognized as a valuable tool for analyzing and optimizing processes, leading to its widespread adoption in industries, and more recently in healthcare. In the context of this thesis, we focus on employing process mining to analyze the daily routine of individuals. In particular, we present a process mining framework designed to identify variations in the execution of behaviors. The idea consists of using sensor systems in the environment to capture the activities of daily living of a person. The data are
then used by a process discovery algorithm to derive a model of personal behavior depicting the typical daily routine. The behavior of the person in the real-world is then continuously perceived by the sensors, and compared with the reference model, in order to identify deviations. The approach finds a possible application in the healthcare field, as a tool to monitor the daily routine of people with neurodegenerative disorders, to provide more effective care. In this case, the identification of variations in the behavior can be the tool for detecting critical health conditions and facilitating interventions.
The thesis comprises seven contributions and additional unpublished novel work. The first contribution consists of a characterization of human behavior to identify the elements that the process model of personal behavior must include. Before delving into the mining of human behavior, two considerations must be addressed. To effectively assess whether existing algorithms can handle processes related to human behavior, it was necessary to construct datasets that accurately reflect the identified characteristics. To this end, we contribute with the implementation of a simulator of human activities, that generates customized synthetic datasets. The output of the simulator is in the form of a stream of sensor events, that are the triggers caused by the movements of the simulated agent in the environment. The second consideration involves the granularity discrepancy between sensor events and the activities in the model. In fact, the model of personal behavior must clearly describe the routine of a person, representing
activities that can be interpreted. On the other hand, the data used to derive the model comes in the form of sensor measurements. For this reason, we have contributed with the implementation of an event abstraction technique that identifies common patterns of events among the data, which can be linked to execution of activities.
Finally, the data can be used for the mining phase. The initial characterization, presented as our primary contribution, contributes to the exploration of process discovery and conformance checking algorithms in the literature. This investigation aims to pinpoint the most suitable algorithms for the analysis of human behavior. In this context, we present two additional contributions. The first acknowledges that human behavior cannot be adequately represented solely through control flow; additional perspectives must be considered. Consequently,
we have implemented a multi dimension mining approach that combines the control flow with statistics to derive and verify an enriched model of personal behavior. The second approach instead, bridges the gap between sensor data and event data at the process level. The approach represents the process model in the form of high level activities. However, during the online conformance checking phase, it processes a stream of sensor events while the model predicates on abstracted activities.
The last part of the thesis focuses on the evaluation and application of the proposed works. We commence with a comparative analysis of the two mining approaches introduced. Following that, we present the application of the multi dimension approach to two real-world datasets. One dataset describes the daily routine of a person, with the aim of studying the evolution of the behavior of the person over time. The second application focuses on the analysis of the routine
of a person with mild cognitive impairment, with the aim of verifying the impact of agitation episodes on the daily routine.
Moreover, the thesis presents a prototype demonstrating the implementation of the presented idea.
Original languageEnglish
PublisherTechnical University of Denmark
Number of pages210
Publication statusPublished - 2023

Fingerprint

Dive into the research topics of 'A Process Mining Framework to Analyze Variability in Human Behavior'. Together they form a unique fingerprint.

Cite this