European Guideline Informed RAG-based GPT-4 Decision Support Tool in Tumor Board Meetings for Breast Cancer Treatment

Breast cancer is the most common cancer among women worldwide. Treatment decision-making in multidisciplinary tumor boards is complex, involving integration of clinical guidelines, patient data, and preferences. We retrospectively evaluated MammaBoardGPT, an few-shot in-context learning and Retrieval Augmented Generation (RAG) enhanced GPT-4 model with European guidelines and five Tumor Board labeled cases, against standard GPT-4 in 25 breast cancer cases discussed at a German hospital. After recursive prompting, MammaBoardGPT achieved complete agreement with tumor board decisions in 84 % of cases, partial agreement in 16 % and no disagreements; standard GPT-4 showed 76 % complete agreement, 20 % partial agreement, and 4 % complete disagreements. A significant difference was observed in MammaBoardGPT's (Stuart-Maxwell P = 0.0048) but not for GPT-4 agreement with tumor board decisions before versus after recursive prompting (Stuart-Maxwell P = 0.135), neither for MammaBoardGPT and GPT-4 in their tumor board decision agreement post-recursive prompting (Stuart-Maxwell P = 0.37). These findings support MammaBoardGPT's potential for tumor board decision support, warranting prospective validation and real-time assessment.