Large language models and synthetic health data: progress and prospects.

Area of Science:

• Health Informatics
• Artificial Intelligence
• Machine Learning

Background:

• High-quality synthetic health data are crucial for advanced analytics due to challenges in acquiring real health data.
• Growing demand for clinical discovery, prediction, and operational excellence necessitates reliable synthetic data solutions.

Purpose of the Study:

• To explore the potential and risks of large language models (LLMs) in synthetic health data generation (SHDG).
• To summarize the current landscape of SHDG methods and identify how LLMs can address existing challenges.

Main Methods:

• Systematic scoping reviews of SHDG domain.
• Analysis of recent LLM methodologies applied to SHDG.
• Investigation of LLM capabilities and limitations.

Main Results:

• Current generative models like Generative Adversarial Networks (GANs) face limitations in SHDG.
• LLMs present promising approaches to mitigate existing challenges in SHDG.
• Six key research directions for LLMs in SHDG were identified: evaluation, adoption, efficiency, generalization, equity, and regulation.

Conclusions:

• LLMs show significant potential but also pose risks within the health domain.
• Further investigation into the advantages and disadvantages of LLMs for SHDG is essential.