RESEARCH · RESEARCH · #1096
HCOE: hyperbolic ontology embeddings map BioBERT into a Poincaré ball for hierarchy-aware clinical concept representation
This arXiv v1 paper introduces Hyperbolic Clinical Ontology Embeddings (HCOE), which map frozen BioBERT embeddings into a Poincaré ball and train with parent-side and child-side ontology-guided contrastive learning plus coarse-to-fine ontology-path aggregation. HCOE is built using ICD codes organized by CCS and ATC medication hierarchies; evaluations report it outperforms baselines on ICD/ATC clinical relation prediction, CCS-to-PheCode hierarchy transfer, and downstream MIMIC‑IV tasks including mortality and readmission prediction, medication recommendation, and rare drug prediction.
KEY POINTS
- This arXiv v1 paper introduces Hyperbolic Clinical Ontology Embeddings (HCOE), which map frozen BioBERT embeddings into a Poincaré ball and train with parent-side and child-side ontology-guided contrastive learning plus coarse-to-fine ontology-path aggregation.
- HCOE is built using ICD codes organized by CCS and ATC medication hierarchies; evaluations report it outperforms baselines on ICD/ATC clinical relation prediction, CCS-to-PheCode hierarchy transfer, and downstream MIMIC‑IV tasks including mortality and readmission prediction, medication recommendation, and rare drug prediction.
- Representing clinical concepts in hyperbolic space can better encode medical code hierarchies, improving clinical relation modeling and downstream EHR prediction tasks.
WHY IT MATTERS
Representing clinical concepts in hyperbolic space can better encode medical code hierarchies, improving clinical relation modeling and downstream EHR prediction tasks.