Advancing Predictive Modeling with Deep Representation Learning on Diverse Clinical Data

Electronic health records (EHRs) provide a wealth of information crucial for enhancing clinical decisions, patient outcomes, and healthcare efficiency. However, leveraging EHR data for predictive modeling involves significant challenges, including managing imbalanced and incomplete data, effectively integrating heterogeneous data types, efficiently capturing temporal and hierarchical structures, and ensuring the interpretability and clinical reliability of predictive outcomes. This dissertation addresses these critical challenges by proposing novel deep representation learning frameworks specifically designed for clinical predictive modeling from diverse EHR data. By integrating sophisticated deep learning techniques, this research demonstrates innovative deep learning modeling approaches and delivers tangible improvements in clinical predictive performance. The dissertation applies these novel frameworks to address three significant, real-world clinical problems: intraoperative hypoxemia prediction, physician burnout prediction, and antimicrobial resistance (AMR) prediction in sepsis patients. For intraoperative hypoxemia prediction, it introduces a Hybrid Inference Network (HiNet), a memory-augmented model designed to robustly predict rare but critical events like hypoxemia, significantly outperforming existing predictive approaches and enhancing clinical decision-making during surgery. For physician burnout prediction, it proposes the Hierarchical Physician Activity Logs (HiPAL) framework, which leverages hierarchical temporal sequence modeling and unsupervised clinical activity embeddings to deliver end-to-end burnout predictions based on raw EHR activity logs, enabling targeted clinical interventions to improve physician well-being and healthcare quality. Finally, to address AMR prediction in sepsis care, the Multimodal Contrastive Clinical Representation Learning (MOCCA) framework is developed, which effectively integrates structured EHR data, clinical notes, and physiological signals through a reliable cross-modal contrastive learning mechanism. The proposed framework shows promise in guiding effective antimicrobial therapies for sepsis patients and improving care quality. Overall, this dissertation contributes to the fields of applied machine learning and healthcare analytics by advancing reliable and clinically impactful predictive models that effectively harness diverse EHR data.