Synthetic Lung-cancer Cohorts Generated by a Large Language Model: Epidemiological Validity Assessment.

Area of Science:

• Medical Informatics
• Oncology
• Artificial Intelligence

Background:

• Large language models (LLMs) are emerging tools in medicine, particularly for clinical reasoning and educational simulations.
• Assessing the validity of AI-generated medical data is crucial for reliable applications.

Purpose of the Study:

• To evaluate the epidemiological plausibility of a synthetic lung cancer cohort generated using ChatGPT-4.0.
• To identify and quantify potential biases in AI-generated epidemiological data.

Main Methods:

• Generated 102 synthetic lung cancer cases in Spanish using structured prompts with demographic, histologic, and molecular data.
• Compared the synthetic cohort descriptively against international datasets (GLOBOCAN 2020, SEER) and meta-analyses.

Main Results:

• The synthetic cohort mirrored general lung cancer patterns but exhibited statistically significant deviations (p < 0.05).
• Early-stage disease and EGFR-positive tumors were overrepresented in the AI cohort.
• Advanced stages, ALK rearrangements, and extreme PD-L1 expression were underrepresented.

Conclusions:

• AI-generated cohorts show generative biases, likely due to training data and model algorithms.
• Despite biases, synthetic cohorts hold potential for educational simulations if methodological transparency is maintained.