A Scoping Review of Synthetic Data Generation by Language Models in Biomedical Research and Application: Data Utility...

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Biomedical research faces challenges with data scarcity, utility, and quality.
• Large Language Models (LLMs) are increasingly adopted for synthetic data generation in this field.
• Existing reviews lack a systematic focus on LLM-driven synthetic data for biomedical applications.

Purpose of the Study:

• To systematically review recent advances in LLM-based synthetic data generation for biomedical and clinical research.
• To analyze how LLMs address data scarcity, utility, and quality across different data modalities.
• To identify current limitations and future directions in this rapidly evolving area.

Main Methods:

• Conducted a scoping review following PRISMA-ScR guidelines.
• Searched major academic databases (PubMed, ACM, Web of Science, Google Scholar) for literature from 2020-2025.
• Included 59 relevant studies focusing on synthetic data generation in biomedical contexts.

Main Results:

• Unstructured text was the predominant data modality (78.0%), followed by tabular (13.6%) and multimodal data (8.4%).
• LLM prompting (74.6%) was the most common generation method, followed by fine-tuning (20.3%).
• Evaluation methods were diverse, with human-in-the-loop assessments (44.1%) being most frequent, followed by intrinsic metrics (27.1%).

Conclusions:

• LLMs offer significant potential for generating synthetic biomedical data, particularly for unstructured text.
• Key barriers include data modality limitations, domain-specific utility, resource accessibility, and the need for standardized evaluation protocols.
• Future research should prioritize developing transparent evaluation frameworks and enhancing accessibility to facilitate widespread adoption in biomedical research.