Bridging LLM Reasoning and UMLS for Spanish Psychiatric Concept Normalization

Resumen

This paper addresses the normalization of clinical concepts in Spanish psychiatric emergency notes, a particularly challenging task due to the use of informal, ambiguous, and context-dependent language, where lexical similarity alone is insufficient to ensure reliable mapping to standardized terminologies. We propose a hybrid architecture that combines semantic retrieval over a UMLS-based knowledge graph with a context-aware reasoning layer guided by clinical entity categories, implemented using LLMs. The method operates on entities extracted through a transformer-based NER model and adapts normalization to six clinical categories. Experiments were conducted on a clinically validated ground truth of 768 entities from 100 psychiatric emergency notes collected at CAULE. Results show substantial improvements over retrieval-only baselines, reaching 70.0% Hit@3 and 73.6% Hit@10 with MiniLM, and 60.9% Hit@3 and 63.3% Hit@10 with Multilingual BERT. Error analysis indicates strong performance in semantically complex categories, while pharmacological entities remain challenging. These findings highlight the effectiveness of combining structured knowledge with contextual reasoning for clinical concept normalization.