Decision-making concordance: A comparative analysis between a multidisciplinary tumor board and a large language model (ChatGPT) in breast cancer management.

Abstract

e13648 Background: Advances in breast cancer management have led to more tailored therapies, increasing the complexity of decision-making. Establishing multidisciplinary tumor boards (MTB) is crucial for optimizing patient care, challenging healthcare teams intellectually. The integration of expertise, deep understanding, and clinical judgment is crucial. Artificial intelligence (AI), particularly through rapid developments and Large Language Models (LLM) offers opportunities to revolutionize healthcare by processing complex ideas and integrating medical evidence into decisions. This study evaluates ChatGPT-4 concordance with a Breast MTB. Methods: This study included 124 patients discussed at our Tumor Board from March 2023 to January 2024, treatment decisions from ChatGPT-4 were compared with those made by the MTB. Results: The median age was 52 years (24-85), with 98% female and 51% premenopausal. In 36% of the patients, the diagnosis followed a screening, while 64% were diagnosed due to symptoms like a palpable nodule or skin changes. 33% of patients meet the NCCN criteria for a family history of cancer. The distribution of clinical stages at diagnosis was 0 (15%), I (35%), II (31%), III (10%), and IV (8%). Most cases were left-sided (52%), followed by right-sided (43%), and bilateral in 5%. The phenotype was Luminal A-like (35%), Luminal B-like HER2 positive (23%), Luminal B-like HER2 negative (15%), and triple negative (10%). The overall concordance between the LLM and MTB rate achieved was 80.65% in patient profiles, including ductal carcinoma in situ. Concordance rates for treatment option indications were high: surgery (96%), radiotherapy (98%), chemotherapy (97%) and endocrine therapy (100%). For additional evaluations (PET/CT or oncogeriatric assessment), the concordance rate was 99%, for indicating genomic signatures was 96%, and for referrals for genetic evaluation, it was 93%. The lowest concordance was observed in type of anti HER2 therapy at 68%. Overall management concordance for patients in stage IV was 50%. Conclusions: AI is seeking ways to integrate into clinical routines by identifying individualized and personalized therapy option for our patients. We explore the concordance and feasibility of employing the LLM ChatGPT-4 in executing one of the most complex tasks in patient care, achieving high rates of concordance. Currently, AI is not capable of replacing medical professionals, but the use of LLM such as ChatGPT as an assistance and support tool for clinicians is likely to expand and evolve. Its capabilities could make it a valuable support tool for therapeutic management in the near future. All authors confirm that this abstract was not thought of or written by ChatGPT.