PrecLLM: A Privacy-Preserving Framework for Efficient Clinical Annotation Extraction from Unstructured EHRs using...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Health

Background:

• Large Language Models (LLMs) excel at text annotation but face privacy and computational hurdles in clinical settings.
• Unstructured Electronic Health Records (EHRs) contain vital data for precision medicine, particularly in oncology.
• Existing LLMs struggle with the scale and privacy demands of clinical data processing.

Purpose of the Study:

• To develop a compact LLM framework (PrecLLM) for efficient and privacy-preserving annotation of clinical notes.
• To enable scalable extraction of crucial information from unstructured EHRs for clinical decision-making.
• To optimize LLM performance for local deployment in resource-constrained and privacy-sensitive healthcare environments.

Main Methods:

• Developed PrecLLM, a novel EHR-specific preprocessing technique to enhance smaller LLMs.
• Integrated regular expressions (regex) and Retrieval-Augmented Generation (RAG) for efficient information extraction.
• Evaluated PrecLLM on private EPIC EHR data (Head and Neck Cancer cohort) and public MIMIC-IV data, comparing against fine-tuned LLMs.

Main Results:

• PrecLLM significantly improved the sensitivity, specificity, and F1 scores of smaller LLMs on EHR tasks.
• Pre-filtering enhanced LLM performance, especially for smaller models on EHR-related tasks.
• PrecLLM demonstrated suitability for privacy-sensitive and resource-constrained healthcare applications.

Conclusions:

• PrecLLM offers optimized LLM performance for secure, efficient, and local healthcare applications.
• The framework addresses key challenges in clinical LLM deployment: privacy, computational feasibility, and clinical applicability.
• Provides practical guidance for leveraging LLMs in healthcare while respecting data security and resource limitations.