Zero-shot Interpretable Phenotyping of Postpartum Hemorrhage Using Large Language Models.

Area of Science:

• Medical informatics
• Natural Language Processing
• 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 training via specific instructions.

Approach:

• Investigated the performance of the Flan-T5 large language model for phenotyping postpartum hemorrhage (PPH) using electronic health record discharge notes.
• Extracted 24 granular concepts related to PPH to develop interpretable phenotypes and subtypes.

Key Points:

• Flan-T5 achieved high fidelity in PPH phenotyping (PPV 0.95), identifying 47% more patients than claims codes.
• The LLM pipeline reliably subtyped PPH, outperforming claims-based methods for common subtypes like uterine atony, abnormal placentation, and obstetric trauma.
• Phenotyping using granular concepts allows for interpretable subtype determination and efficient algorithm updates.

Conclusions:

• LLM-based phenotyping offers a rapid, reliable method for identifying PPH and its subtypes directly from clinical notes.
• This approach bypasses the need for manually annotated training data, enabling broad application across various clinical use cases.
• The interpretability of LLM-derived phenotypes facilitates clinical understanding and adaptation to evolving medical guidelines.