Modelling and predicting medical outcomes for intensive care patients via network mechanisms and machine learning

Abstract

In the intensive care unit (ICU), where patients receive critical life support, providing effective care is challenging due to the diverse needs, medical backgrounds, and demographics of patients. To aid in decision-making, patient data is continuously monitored, offering a wealth of information. While machine learning has been explored to assist in ICU decision-making, challenges like data quality and patient diversity hinder practical implementation.

This thesis introduces a novel approach that combines network representation of clinical data with community detection to enhance existing models and address ICU care challenges. The method is applied to two ICU tasks: identifying patient multimorbidity profiles and predicting in-hospital mortality. Through these studies, the limitations of existing methods are highlighted, and the benefits of the proposed approach are demonstrated.

The results showcase several advantages: the robust network representation mitigates the impact of missing, captures heterogeneity factors like ethnicity and age on mortality, offers flexibility to adapt to varying availability of data, eliminates complex parameter selection, and provides a visually intuitive model delivery.

This work represents progress in leveraging networks and machine learning to support decision-making in the complex ICU environment. Future research directions include refining data representation, exploring network reconstruction for mortality prediction, and enhancing model interpretability with medical knowledge.