Two Directions for Clinical Data Generation with Large Language Models: Data-to-Label and Label-to-Data

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Clinical text mining is challenging due to scarce, sensitive, and imbalanced medical data.
• Detecting Alzheimer's Disease (AD) signs and symptoms from electronic health records (EHRs) requires specialized expertise.

Purpose of the Study:

• To investigate the potential of Large Language Models (LLMs) in augmenting clinical data for AD detection.
• To develop and evaluate methods for generating synthetic clinical data for AD sign and symptom detection.

Main Methods:

• Creation of a novel pragmatic taxonomy for AD sign and symptom progression.
• Generation of three datasets: gold (expert-annotated EHRs), silver (data-to-label method), and bronze (label-to-data method).
• Training a system to detect AD-related signs and symptoms using the generated datasets.

Main Results:

• The system trained with silver and bronze datasets showed improved performance compared to using only the gold dataset.
• LLMs can generate synthetic clinical data for complex tasks by incorporating expert knowledge.
• The label-to-data method produced high-quality, privacy-preserving synthetic datasets.

Conclusions:

• LLMs are effective in generating synthetic clinical data to enhance AD detection from EHRs.
• Synthetic data generation, particularly using the label-to-data method, offers a viable solution for data scarcity and privacy concerns in clinical text mining.
• This approach can improve the performance of AI systems in identifying complex medical conditions.