Abstract

Electronic health records (EHRs) are inherently irregularly sampled, posing significant challenges for conventional time-series models. In this work, we introduce HealthODE—a novel framework that leverages neural discretized ordinary differential equations (ODEs) to learn robust representations from irregular health data. By integrating a decay-gated attention mechanism and rotary positional encoding, HealthODE adaptively filters irrelevant historical data while accurately capturing continuous dynamics. Our approach supports both interpolations within observed intervals and extrapolation beyond them, enabling zero-shot forecasting for a range of clinical tasks such as diagnostic prediction, drug usage estimation, and phenotype classification. Empirical evaluations demonstrate that HealthODE not only improves forecasting accuracy but also provides interpretable insights into patient risk trajectories, making it a promising tool for advanced healthcare analytics.

Authors

Shrutibahen Patel
Fairleigh Dickinson University, United States of America

Keywords

HealthODE, Historical data, Zero-Shot Forecasting, Clinical Task