EHR Forecasting via Implicit Neural ODEs

Abstract

The widespread adoption of Electronic Health Record (EHR) systems has generated vast repositories of valuable clinical data. Utilizing this information enables the development of predictive methods for proactive healthcare, enhanced medical services, and optimized resource allocation.This study employs the MIMIC-IV dataset to address challenges in generating actionable insights from longitudinal EHR data. We introduce a novel model based on Implicit Neural Ordinary Differential Equations (Neural ODEs), focusing on predicting Length of Stay (LoS) and In-Hospital Mortality (IHM). The model first encodes multimodal clinical codes into a low-dimensional space via a clinical embedding module, mitigating noise from heterogeneous data. An implicit Neural ODE network then captures the temporal dynamics of vital signs and lab results for efficient continuous-time modeling，The key experimental results show that the AUC of the proposed model in the prediction of hospital stay reaches 0.81, and the accuracy reaches 97%, which is 12% and 9% higher than that of the baseline model, respectively. In terms of mortality prediction, it also showed advantages, with AUC reaching 0.785 and mean square error decreasing from 0.0084 to 0.0036, a decrease of about 39.5%, which was significantly better than the prediction effect of the baseline model.