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

Area of Science:

• Computational biology
• Medical informatics
• Genomics

Background:

• Structured phenotypes are crucial for Mendelian disorder diagnosis and genetic research.
• Electronic health records (EHRs) contain vast phenotypic data, but it is largely unstructured.
• Existing automated phenotyping methods struggle with semantic variability and contextual information in clinical narratives.

Purpose of the Study:

• To develop an advanced automated phenotyping framework addressing limitations of current deep learning models.
• To improve the extraction of standardized Human Phenotype Ontology (HPO) identifiers from unstructured clinical text.
• To enhance the utility of EHR data for genetic studies and clinical decision support.

Main Methods:

• Proposed LEAP, a two-stage framework combining a large language model (LLM) for phenotype extraction and a fine-tuned sentence-transformer for HPO mapping.
• The LLM component handles long clinical narratives without text chunking.
• The sentence-transformer model is trained on over 5.3 million instances for accurate and deterministic HPO identifier generation.

Main Results:

• LEAP demonstrated significant relative improvements in precision (19.68%-412.68%) and F1 score (44.14%-298.77%) on a real-world EHR test set compared to existing tools.
• Achieved robust performance on external benchmarks, validating its generalizability.
• The framework ensures the output of valid and deterministic HPO identifiers, overcoming LLM limitations.

Conclusions:

• LEAP offers a robust and accurate solution for automated phenotyping from unstructured EHR data.
• The framework enhances the standardization and usability of phenotypic data for downstream analyses, including gene prioritization.
• LEAP represents a significant advancement in leveraging AI for clinical informatics and precision medicine.