Bridging clinical narratives and structured phenotypes with large language models and sentence transformers

Structured phenotypes are important for Mendelian disorder diagnosis, gene-phenotype association studies, and standardized phenotypic data sharing. Although electronic health records (EHRs) contain abundant phenotypic information, much of it is unstructured. Early automated phenotyping methods are rule-based, limiting their ability to capture semantic variability and contextual information. Recent deep learning approaches, including BERT-based models and large language models (LLMs), improve semantic understanding but still face key limitations. BERT-based methods are constrained by limited context windows, requiring text chunking and aggregation for long clinical narratives, while LLMs that directly generate Human Phenotype Ontology (HPO) identifiers may produce non-existent identifiers. To address these challenges, we propose LLM-Enhanced Automated Phenotyping (LEAP), a two-stage framework that integrates an LLM for free-text phenotype extraction with a sentence-transformer model fine-tuned on a large-scale dataset of 5,330,557 instances for HPO mapping. This design handles long inputs while ensuring valid and deterministic HPO identifier outputs. On a real-world EHR test set, LEAP achieves relative improvements of 19.68%-412.68% in precision and 44.14%-298.77% in F1 score compared with existing tools, while maintaining robust performance on external benchmarks. LEAP can be integrated with gene prioritization tools to provide standardized phenotype inputs for downstream analyses. LEAP is available at phenogemini.org/extract.