COMBINED RISK MODELING AND SUBTYPING IN INTENSIVE CARE UNITS

Abstract

In this thesis, subtype-aware risk modeling approaches are developed for clinical decision support systems in intensive care units (ICUs). Traditional risk models often fail to account for the heterogeneous subpopulations within diseases, leading to inadequate accuracy in risk prediction. To address this, a computational design science paradigm is adopted. Theoretical analysis identifies the importance of class separability in improving risk prediction performance. Based on this, a novel k-means-based classification algorithm, Classification Aware Clustering (CAC), and its Deep Neural variant, DeepCAC, are proposed to effectively identify subtypes. Additionally, a subtype-aware risk modeling approach called ExpertNet is developed, utilizing deep neural networks to model heterogeneity and cluster-specific classifiers. Experiments on real ICU data demonstrate that ExpertNet significantly outperforms existing approaches in predicting complications like Sepsis and Acute Respiratory Distress Syndrome (ARDS), offering valuable insights for personalized care strategies.