SynCABEL: Synthetic Contextualized Augmentation for Biomedical Entity Linking

TL;DR

SynCABEL tackles the data bottleneck in biomedical entity linking by using large language models to generate context-rich synthetic examples for all candidate concepts in a KB, enabling effective training of decoder-only BEL models. The approach achieves state-of-the-art performance on English, French, and Spanish BEL benchmarks and demonstrates data efficiency by matching fully supervised results with substantially fewer hand-annotated data. A novel LLM-as-a-judge protocol assesses clinical validity beyond exact label matches, revealing improved semantic alignment and practical usefulness. The work highlights the value of comprehensive synthetic augmentation, guided inference, and adaptive concept representations for scalable, multilingual BEL with real-world applicability.

Abstract

We present SynCABEL (Synthetic Contextualized Augmentation for Biomedical Entity Linking), a framework that addresses a central bottleneck in supervised biomedical entity linking (BEL): the scarcity of expert-annotated training data. SynCABEL leverages large language models to generate context-rich synthetic training examples for all candidate concepts in a target knowledge base, providing broad supervision without manual annotation. We demonstrate that SynCABEL, when combined with decoder-only models and guided inference establish new state-of-the-art results across three widely used multilingual benchmarks: MedMentions for English, QUAERO for French, and SPACCC for Spanish. Evaluating data efficiency, we show that SynCABEL reaches the performance of full human supervision using up to 60% less annotated data, substantially reducing reliance on labor-intensive and costly expert labeling. Finally, acknowledging that standard evaluation based on exact code matching often underestimates clinically valid predictions due to ontology redundancy, we introduce an LLM-as-a-judge protocol. This analysis reveals that SynCABEL significantly improves the rate of clinically valid predictions. Our synthetic datasets, models, and code are released to support reproducibility and future research.

Figures (5)

• Figure 1: Illustration of biomedical entity linking annotation scarcity: a context-free model fails without context, a supervised context-aware model trained on human annotations fails on unseen concepts, while a SynCABEL-augmented model leverages synthetic data to recover the correct concepts.
• Figure 2: Example of the disambiguation process used to select a natural representation $e_m$ = "Fluid Discharge" for a concept $e$ = "C0012621" and a mention $m$ = "discharge".
• Figure 3: Overview of SynCABEL. LLM: large language model
• Figure 4: LLM-as-a-judge evaluation of predicted vs. gold concepts using four clinical-relation classes: Correct, Broad, Narrow, and No relation.
• Figure 5: Data Efficiency Analysis. Exact-match Recall@1 performance as a function of human training data availability. The dashed line marks the performance ceiling of the standard baseline trained on human-annotated data. Stars ($\star$) indicate where the SynCABEL-augmented model (Blue) matches or surpasses this ceiling. 0% implies training on synthetic data only. Error bars denote 95% bootstrap CIs.