Clinical Information Extraction with Large Language Models: A Case Study on Organ Procurement.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Healthcare Data Extraction

Background:

• Large language models (LLMs) show promise for clinical text analysis.
• Existing LLM applications often overlook extracting numeric clinical data (e.g., lab tests, vital signs).

Purpose of the Study:

• To evaluate LLMs' capability in extracting numeric data from clinical text within organ procurement.
• To develop and validate an LLM-based approach for numeric data extraction in this domain.

Main Methods:

• Developed an LLM approach incorporating a specific prompting strategy for numeric extraction.
• Implemented novel heuristics to mitigate hallucination in extracted data.
• Validated the approach on 298 hand-annotated clinical notes.

Main Results:

• The LLM approach achieved high accuracy, precision, and recall in extracting numeric data.
• Demonstrated the approach's utility on a large dataset of 43,719 notes from 14,342 potential organ donors.

Conclusions:

• LLMs are effective for extracting numeric clinical data in organ procurement settings.
• This method enhances downstream data analysis and supports making organ procurement data accessible for research.