Brief

Researchers have developed new methods for electronic health record (EHR) foundation models to better utilize the hierarchical structure of ICD diagnosis codes. Current models treat these codes as flat tokens, ignoring their inherent relationships. This work explores augmenting BERT-style transformers with hierarchical tokens and incorporating hierarchy into graph-based code representations. Experiments on MIMIC-IV and eICU datasets demonstrate that explicitly encoding ICD hierarchy improves downstream prediction accuracy and cross-dataset transferability, with the most effective hierarchy level varying by task and model. AI

Researchers explored the effectiveness of generalistic versus specific embeddings for semantic search in clinical coding across non-English languages. They found that fine-tuning a Spanish biomedical encoder with LLM-generated synthetic data significantly improved performance in languages like Spanish, Catalan, French, and Portuguese. This approach, involving a bi-encoder and a cross-encoder reranker, even surpassed existing English-based models on certain metrics without English biomedical pretraining. AI

Researchers have developed RAG-Coding, a novel method that uses four large language model (LLM) agents to improve the accuracy of automated medical coding for ICD-10-CM. This approach grounds the LLMs' decisions in external knowledge sources like official coding guidelines and tabular lists, enhancing clinical compliance. In evaluations on the MDACE dataset, RAG-Coding demonstrated significant improvements over existing LLM-based baselines, achieving higher micro-F1 and macro-F1 scores. The study also introduced an updated dataset, MDACE-2025, which incorporates the latest 2025 ICD-10-CM guidelines for more precise evaluation. AI