Introduction: Multidisciplinary tumor boards (MTBs) are the standard for oncologic decision-making but require substantial time and personnel resources. Large language models (LLMs) may support MTBs by generating structured, guideline-based recommendations. We evaluated concordance between MTB decisions and LLM recommendations in newly diagnosed cholangiocellular adenocarcinoma (CCA) and hepatocellular carcinoma (HCC). Method: In this retrospective single-center study, MTB protocols from 2022–2023 were screened. After exclusions, 50 CCA and 50 HCC cases were analyzed. Institutional MTB recommendations were compared with outputs from ChatGPT and Claude using identical structured clinical summaries available at the time of MTB presentation. Agreement was assessed using Cohen’s kappa, and correlation using Spearman’s rank coefficient. Results: For CCA, exact concordance between MTB and ChatGPT was 80%, with substantial agreement (k = 0.688) and strong correlation (r = 0.725; both p < 0.001). Concordance with Claude was at 56%. For HCC, concordance between MTB and ChatGPT was 66%, with substantial agreement (k = 0.604) and moderate correlation (r = 0.484; both p < 0.001). Concordance with Claude was only 38%, with fair agreement (k = 0.314, p < 0.001) and weak correlation (r = 0.086, p = 0.551). Conclusions: LLM-derived recommendations varied markedly by model. ChatGPT demonstrated substantial concordance, whereas Claude showed limited agreement. MTBs appeared more individualized, while LLM outputs were more guideline-oriented. These findings highlight the potential role of LLMs as supportive tools for structured clinical reasoning and guideline adherence. Prospective multicenter studies should evaluate real-time LLM integration into MTB workflows, efficiency, decision quality, and patient outcomes safely.
Palzer et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: