Large Language Models as Decision Support in Breast Cancer Management: A Comparison with Multidisciplinary Team...

Background/Objectives: Breast cancer management relies on multidisciplinary team (MDT) decisions that integrate clinical, radiological, pathological, and patient-related factors. Large language models (LLMs) may support such decisions, but evidence based on real-world cases remains limited. Methods: We evaluated the agreement between ChatGPT-5.5 (OpenAI, San Francisco, CA, USA) and real MDT management of patients with invasive breast cancer. In this single-center, retrospective, proof-of-concept study, standardized preoperative clinical vignettes for 86 patients with invasive breast carcinoma were independently evaluated by GPT-5.5 (frozen prompt, Medium reasoning mode, Web Search enabled). AI-generated recommendations were compared with the MDT reference standard across eight therapeutic domains (four surgical, four non-surgical), yielding 688 individual decisions. Agreement was summarized descriptively and with Cohen's κ. Results: GPT-5.5 produced complete recommendations for all vignettes. Overall concordance was 90.8% (625/688). Agreement was highest for endocrine therapy, radiotherapy, and targeted therapy (100% each) and for axillary surgery (93.0%, κ  =  0.86), and lowest for breast reconstruction (80.2%), oncoplastic surgery (82.6%) and neoadjuvant chemotherapy (83.7%). Surgical concordance was 85.8% (295/344). Discordances occurred predominantly in individualized surgical decisions; none were observed for endocrine therapy, radiotherapy, or targeted therapy. Conclusions: GPT-5.5 showed high overall concordance with real MDT decisions, particularly for guideline-driven systemic therapy, but was less reliable for individualized surgical planning. These findings should be regarded as exploratory and cannot be assumed to be generalizable across institutions or clinical settings. LLMs should complement, not replace, multidisciplinary expertise; prospective multicenter validation is warranted.