Exploring Detection Methods for Synthetic Medical Datasets Created With a Large Language Model.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Data Science

Background:

• Large language models (LLMs), such as Generative Pre-trained Transformer 4 (GPT-4), have demonstrated the capability to generate synthetic medical datasets.
• These synthetic datasets can be designed to support fabricated scientific evidence, raising concerns about research integrity.

Purpose of the Study:

• To investigate the statistical patterns indicative of data fabrication by LLMs.
• To explore methods for improving synthetic datasets to evade authenticity checks and enhance their reliability.

Main Methods:

• Synthetic datasets were generated for three fictional clinical studies using GPT-4o and a custom GPT model.
• Forensic analysis was conducted on both unrefined and refined datasets to identify statistical anomalies and unrealistic clinical records.

Main Results:

• Initial forensic analysis revealed numerous fabrication marks (33.9%) in unrefined datasets, including gender-name mismatches, weekend visit dates, and age calculation errors.
• Refined datasets showed a significant reduction in fabrication signs (4.6%), with four datasets passing forensic analysis.
• However, some refined datasets still exhibited suspicious characteristics, such as unusual distribution shapes.

Conclusions:

• Advanced custom GPT models can generate synthetic data that may pass forensic scrutiny.
• The potential for AI to fabricate authentic-seeming datasets necessitates robust detection methods to maintain scientific validity.