Voorspelling van ongewenste voorvallen op de intensive care met behulp van machine learning

The digitalization of healthcare in recent years has led to a constantly increasing amount of accessible data generated in clinical settings, paving the way towards the development of data-driven decision support applications. In intensive care units, where patients are continuously monitored due to their critical condition and clinical decisions must often be made under time pressure, machine learning models have the potential to provide personalized risk assessments to support clinicians in their workflow. However, the use of machine learning in this setting raises methodological and ethical challenges. Among them, interpretability represents an important condition for clinical integration, as healthcare professionals require transparency not only to trust and understand model outputs but also to ensure that predictions align with medical reasoning and do not hinder patient safety. This PhD thesis concerns the development of machine learning approaches designed to handle the different aspects of the complexity of clinical data. We focus on modelling time-to-event outcomes, common in healthcare scenarios, using methodologies that allow for post hoc or intrinsic interpretability while maintaining predictive accuracy. The contributions of this work span from algorithmic development to clinical applications, with an emphasis on problems relevant to the management of critical care patients. The research is structured into four objectives. The first objective focuses on intubation for mechanical ventilation. First, we investigate the use of survival models that incorporate cured fraction information to identify patients at risk of intubation for mechanical ventilation after intensive care unit admission. Second, we use an outcome-aware clustering approach to characterize the distinct phenotypes of mechanically ventilated patients prior to extubation. These phenotypes capture variability in the physiological state of the patients and clinical time-to-event outcomes, and are designed to predict extubation failure. The second objective focuses on the development of a model for early prediction of delirium, a common but often overlooked complication in critical care. Our approach uses routinely available intensive care unit data to implement a predictive model aimed at improving resource allocation for targeted prevention. The third objective involves the use of sequential pattern mining applied to intensive care unit monitor data, where our aim is to extract clinically relevant signals and alarm patterns from unvalidated logs. This work offers insights into signals and alarm sequences that precede adverse events, to pave the way to the development of smart alarm systems. In the last objective, we explore a reframing of multi-event survival analysis as a regression problem, enabling the use of off the shelf machine learning models in predicting multiple events simultaneously. These contributions are primarily evaluated on real-world critical care data collected from a secondary care hospital in Belgium, complemented by additional synthetic and real-world datasets. By aligning predictive performance with explainability, this work contributes to the broader effort to integrate machine learning into critical care medicine in a safe, transparent, and clinically meaningful way.