HNMP-EHR: Ontology-aligned Representation Learning on Electronic Health Records for Concurrent Prediction

S Mehryar, Michel Dumontier

Keywords: EHR, Graph Representation Learning, Hyperbolic Neural Message Passing, Concurrent Prediction

Abstract: The growing availability of electronic health records (EHRs) presents ever more opportunities for data-driven applications in healthcare. However, these records often suffer from fragmentation and structural inconsistencies, where key contextual elements are frequently missing. In this work, we present a system that enhances clinical predictions based on representation learning through ontology alignment. To address data gaps, we introduce a dual-space strategy that guides the learning process over heterogeneous entities. To address structural conformity, we introduce a hyperbolic message passing algorithm that allows the integration of hierarchical information. Our experiments, focused on cardiovascular disease (CVD) patients, demonstrate that the system improves performance across multiple tasks, including lab test, diagnosis, medication, and mortality prediction, performed concurrently. The results highlight the potential of ontology-aligned representation learning for high-precision and multi-purpose decision-making in healthcare systems.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 1579