HazardFlow: Enhancing Health Status Representations via Score-based Energy Modeling

Learning effective health status representations from Electronic Health Records (EHRs) is essential for accurate health risk prediction. While data-driven models have advanced this field, most existing approaches rely on heuristic patterns and often struggle to capture subtle risk dynamics, specifically under imbalanced label distributions. In this work, we propose a novel score-based energy modeling paradigm to enhance health status representation by incorporating a mathematically grounded formulation of health risk. We theoretically show that the hazard function, a canonical quantification of instantaneous risk, is mathematically proportional to the score function of the health status representation, providing a principled mechanism for modeling risk dynamics. Building on this insight, we develop HazardFlow, a pluggable module that explicitly estimates the hazard function and enhances health status representations with hazard escalation and reduction mechanisms. HazardFlow is designed for seamless integration into diverse health risk prediction models, including both uni-modal and multi-modal backbones. Extensive experiments on the MIMIC-III, MIMIC-IV, and MIMIC-CXR datasets demonstrate that HazardFlow consistently improves performance across diverse health risk prediction tasks, validating its generality, robustness, and theoretical soundness.