HCOE: Hyperbolic Clinical Ontology Embeddings for Biomedical LMs

What happened

Two research teams released two new embedding models respectively: HCOE, built to address the painpoint that existing biomedical language models do not explicitly preserve medical code hierarchies. HCOE maps frozen BioBERT embeddings to a Poincaré ball, trained on hierarchical ICD disease codes and ATC drug classification data, and achieves state-of-the-art performance across multiple clinical tasks on the MIMIC-IV dataset; and SEA-CLIP-Tiny, built to address weak support for text-image embedding in Southeast Asian languages. The model has fewer than 50M parameters, is trained via knowledge distillation, supports 7 Southeast Asian languages, and has 38.4% fewer parameters than MobileCLIP2, with an average R@10 improvement of 12.1 percentage points.

Key facts

Publication date

September 28, 2026

Publication channel

arXiv preprint platform

First released product

HCOE hyperbolic clinical ontology embedding model

Second released product

SEA-CLIP-Tiny lightweight multilingual text-image embedding model

HCOE benchmark performance

Achieves state-of-the-art performance across multiple clinical tasks on the MIMIC-IV dataset

SEA-CLIP-Tiny benchmark performance

Has 38.4% fewer parameters than MobileCLIP2, with an average R@10 improvement of 12.1 percentage points

Background

Existing biomedical AI models did not explicitly preserve medical code hierarchies, and high-efficiency multilingual text-image embedding models for low-compute Southeast Asian language scenarios were lacking. Current solutions suffer from poor adaptability, high deployment costs, and insufficient accuracy.

Why it matters

For the broader industry, these two solutions break through technical bottlenecks in healthcare AI and Southeast Asian regional multimodal applications, validating the value of hierarchy-aware and region-adapted small model training. For developers, they reduce the training, adaptation, and deployment costs for these two application domains. For end users, they enable more accurate clinical decision support and a more local-language-aligned multimodal service experience.

What to watch

Future tracking should focus on the open-source progress of these two models, real-world deployment test results, and industry reuse of their training methodologies.