Single-session agreement of ChatGPT and Gemini treatment recommendations with multidisciplinary tumor board decisions...

To evaluate the single-session agreement between treatment recommendations generated by ChatGPT and Gemini and real-world decisions made by a thoracic oncology multidisciplinary tumor board (MDTB).

This retrospective observational study included 102 thoracic oncology cases discussed at a university hospital MDTB between June and December 2025. Cases were categorized as primary lung cancer, pulmonary mass without pathological diagnosis, esophageal cancer, or pulmonary metastasis. Standardized anonymized case summaries, including clinical, radiological, pathological, and laboratory data, were presented to ChatGPT and Gemini. Each model was asked to provide a single tumor board-like management recommendation. Each case was submitted once to each model in a separate new chat session, and the initial output was used for analysis. Recommendations were classified into predefined categories and compared with final MDTB decisions. Agreement was assessed using percentage agreement and Cohen's kappa coefficient.

In this single-session evaluation, overall agreement with MDTB decisions was 52.0% for ChatGPT and 56.9% for Gemini. Both models showed moderate agreement, with slightly higher agreement for Gemini than ChatGPT (κ = 0.487 vs. κ = 0.432; p < 0.001). Agreement was higher in standardized scenarios, such as primary lung cancer and esophageal cancer, but lower in pulmonary masses without pathological diagnosis and pulmonary metastases. No significant difference was found between the models in overall concordance with MDTB decisions (p = 0.359).

The initial outputs of ChatGPT and Gemini showed moderate agreement with MDTB decisions in this single-session evaluation. LLMs may support guideline-based thoracic oncology decision-making but remain limited in complex, patient-specific scenarios and should not replace multidisciplinary clinical judgment. These findings should be interpreted as concordance with real-world MDTB decisions rather than evidence of stable model-level performance or clinical accuracy.