Concordance Between a Generative Artificial Intelligence Model and a Hepatobiliary Multidisciplinary Team in Hepatocellular Carcinoma Management: A Retrospective Study.

Abstract

The management of hepatocellular carcinoma (HCC) relies on decisions formulated by a multidisciplinary team (MDT). Generative artificial intelligence (AI) models are increasingly explored as decision-support tools. This study aimed to assess the level of agreement between AI-generated recommendation and expert MDT decisions in HCC management. A retrospective cohort of 100 HCC patients discussed at a hepatobiliary MDT between January 2023 and February 2025 was retrospectively included. Anonymized case summaries, excluding MDT conclusion, were submitted to a large language AI model (ChatGPT-4). Recommendations were categorized using the following four-level progressive concordance scale: (1) treatment aim (curative vs palliative), (2) treatment extent (locoregional vs systemic), (3) treatment modality (e.g., resection, ablation, embolization), and (4) specific technique. Concordance levels 1 and 2 were classified as low concordance (LC), whereas levels 3 and 4 were classified as high concordance (HC). In 81 % of the cases, HC between AI-generated and MDT recommendations was observed, with 53 % showing complete agreement on treatment technique. In 19 % of the cases, LC was identified and associated with older patient age (p = 0.001). In most discordant cases, discrepancies reflected the MDT's preference for more aggressive, tailored interventions. In HCC management, AI demonstrated a high level of agreement with expert MDT decisions, suggesting its potential role as a complementary decision-support tool. However, limitations persist for elderly patients and borderline clinical scenarios, in which individualized human judgment remains essential.