Process Mining in Healthcare: Maintenance of Data and Models

This thesis investigates the complexities associated with the application of Process Mining (PM) and Machine Learning (ML) in the healthcare domain, with a specific focus on the long-term maintenance of data and models. Despite the transformative potential of these technologies in optimizing clinical pathways and operational efficiency, their adoption is frequently hindered by the high heterogeneity of medical data, the non-linear nature of clinical processes, and the performance degradation of predictive models over time, commonly known as concept drift. To address these challenges, this research proposes an integrated framework that encompasses the entire data lifecycle, ranging from automated acquisition and semantic standardization to robust process discovery and adaptive predictive modelling. The primary contribution of this work lies in the development of a pipeline for the semantic reconciliation of heterogeneous and evolving healthcare data, complemented by a process-aware methodology designed to extract complex temporal and behavioral features from care pathways. Furthermore, the proposal introduces a comprehensive governance and maintenance layer dedicated to continuous model monitoring, drift detection, and human-in-the-loop validation. Experimental results, validated on real-world clinical datasets, demonstrate that integrating a process- oriented perspective significantly enhances the accuracy, interpretability, and resilience of predictive analytics within healthcare environments. Keywords: Process Mining, Healthcare Analytics, Machine Learning Maintenance, Concept Drift Detection, Clinical Pathways, Data Governance, Predictive Modelling, Evidence-based Medicine