Knowledge-Infused Prompting: Assessing and Advancing Clinical Text Data Generation with Large Language Models.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Clinical natural language processing (NLP) is hindered by complex medical terminology and context.
• Large language models (LLMs) show potential but face privacy and resource limitations for direct clinical application.

Purpose of the Study:

• To develop a resource-efficient method for generating synthetic clinical text using LLMs for clinical NLP tasks.
• To address privacy concerns and resource constraints associated with direct LLM deployment in healthcare.

Main Methods:

• Proposed ClinGen, an innovative approach integrating clinical knowledge extraction and context-informed LLM prompting.
• Utilized domain-specific knowledge graphs and LLMs to guide the generation of clinical topics and writing styles.
• Employed a resource-efficient strategy for synthetic data generation.

Main Results:

• ClinGen demonstrated consistent performance enhancements across 8 clinical NLP tasks and 18 datasets, averaging 7.7%-8.7%.
• The generated synthetic data effectively aligned with the distribution of real clinical datasets.
• ClinGen enriched the diversity of training instances, improving model generalization.

Conclusions:

• ClinGen offers an effective and efficient solution for synthetic clinical text generation.
• The approach successfully mitigates privacy and resource challenges in clinical NLP.
• ClinGen significantly improves performance and data diversity for various clinical NLP applications.