Merve TOKOÇİN, Turan PEHLİVAN, Selçuk CİN, Bülent TOKSÖZ, Onur TOKOÇİN, Eda CİNGÖZ, Nigar ERKOÇ, Aynur ÖZEN, Nida SÜNNETÇİ ARIKAN, Şahin BEDİR, Atilla ÇELİK
European Journal of Breast Health - 2026;22(2):184-189
Objective: Multidisciplinary teams (MDTs) are essential for optimizing breast cancer treatment, yet the role of general-purpose artificial intelligence (AI), such as ChatGPT, in supporting these teams remains underexplored. This study compared ChatGPT versions 3.5 and 4 with a hospital-based MDT in making treatment and follow-up recommendations, using St. Gallen, European Society for Medical Oncology, National Comprehensive Cancer Network, and American Society of Clinical Oncology guidelines as a reference. Materials and Methods: A retrospective analysis of 100 consecutive breast cancer patients diagnosed between January 2023 and January 2024 at a training hospital in İstanbul, Türkiye, was conducted. The MDT provided consensus-based recommendations, while anonymized patient data were processed by ChatGPT using English prompts based on guideline summaries. Two experienced breast surgeons independently rated recommendation appropriateness on a five-point scale post-treatment, focusing on clinical outcomes, with agreement assessed using weighted Cohen's kappa across cancer stage, molecular subtype, and proliferation index. Results: ChatGPT-4 (with a knowledge cut-off of March 2023) demonstrated substantial agreement with the MDT for primary treatments (weighted kappa = 0.712), whereas ChatGPT-3.5 showed moderate agreement (kappa = 0.600). Agreement for additional recommendations, such as genetic counseling, was lower (GPT-4: kappa = 0.398; GPT-3.5: kappa = 0.302), with better performance in early-stage and less aggressive subtypes compared to advanced or aggressive cases. Discrepancies were noted in complex or aggressive cases. Conclusion: The study suggests ChatGPT, particularly version 4, may serve as a supportive tool for breast cancer teams, especially in early-stage cases, though clinical expertise remains vital for complex scenarios, warranting further research to refine AI integration.
