Zero-shot interpretable phenotyping of postpartum hemorrhage using large language models.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Clinical Phenotyping

Background:

• Accurate patient phenotyping is crucial in medicine but often limited by the need for extensive annotated data.
• Large language models (LLMs) show promise for adapting to new tasks with minimal or no task-specific training.

Purpose of the Study:

• To evaluate the performance of a publicly available LLM, Flan-T5, for phenotyping postpartum hemorrhage (PPH) using electronic health record (EHR) discharge notes.
• To assess the LLM's ability to identify granular concepts for developing interpretable PPH subtypes.

Main Methods:

• Utilized Flan-T5, a large language model, to analyze 271,081 EHR discharge notes for PPH phenotyping.
• Extracted 24 granular concepts related to PPH and compared LLM performance against standard claims codes for patient identification and subtype analysis.

Main Results:

• Flan-T5 achieved high fidelity in PPH phenotyping (PPV of 0.95), identifying 47% more patients than claims codes.
• The LLM pipeline demonstrated superior performance in subtyping PPH, particularly for uterine atony, abnormal placentation, and obstetric trauma.
• The approach allows for interpretable phenotyping and efficient algorithm updates as clinical guidelines evolve.

Conclusions:

• LLMs like Flan-T5 offer a powerful, rapid, and interpretable method for clinical phenotyping using EHR notes.
• This LLM-based approach significantly improves PPH identification and subtyping compared to traditional claims-based methods.
• The ability to phenotype without manually annotated data opens new avenues for clinical research and precision medicine.