HypEHR: Hyperbolic Model for Efficient EHR Question Answering

Researchers have introduced HypEHR, a streamlined Lorentzian framework designed for answering questions related to electronic health records by embedding codes, visits, and inquiries within hyperbolic space. This model utilizes hyperbolic geometry to effectively represent the hierarchical nature of clinical data, inspired by findings that suggest medical ontologies and patient pathways reflect this geometry. HypEHR employs geometry-consistent cross-attention along with type-specific pointer heads for query responses. It undergoes pretraining focused on predicting next-visit diagnoses and incorporates hierarchy-aware regularization to ensure alignment with the ICD ontology. In evaluations on two MIMIC-IV-based EHR-QA benchmarks, HypEHR nearly matches LLM-based approaches while requiring significantly fewer parameters. The source code is publicly accessible.