The imitation game: large language models versus multidisciplinary tumor boards: benchmarking AI against 21 sarcoma centers from the ring trial.

PURPOSE

The study aims to compare the treatment recommendations generated by four leading large language models (LLMs) with those from 21 sarcoma centers' multidisciplinary tumor boards (MTBs) of the sarcoma ring trial in managing complex soft tissue sarcoma (STS) cases.

METHODS

We simulated STS-MTBs using four LLMs-Llama 3.2-vison: 90b, Claude 3.5 Sonnet, DeepSeek-R1, and OpenAI-o1 across five anonymized STS cases from the sarcoma ring trial. Each model was queried 21 times per case using a standardized prompt, and the responses were compared with human MTBs in terms of intra-model consistency, treatment recommendation alignment, alternative recommendations, and source citation.

RESULTS

LLMs demonstrated high inter-model and intra-model consistency in only 20% of cases, and their recommendations aligned with human consensus in only 20-60% of cases. The model with the highest concordance with the most common MTB recommendation, Claude 3.5 Sonnet, aligned with experts in only 60% of cases. Notably, the recommendations across MTBs were highly heterogenous, contextualizing the variable LLM performance. Discrepancies were particularly notable, where common human recommendations were often absent in LLM outputs. Additionally, the sources for the recommendation rationale of LLMs were clearly derived from the German S3 sarcoma guidelines in only 24.8% to 55.2% of the responses. LLMs occasionally suggested potentially harmful information were also observed in alternative recommendations.

CONCLUSIONS

Despite the considerable heterogeneity observed in MTB recommendations, the significant discrepancies and potentially harmful recommendations highlight current AI tools' limitations, underscoring that referral to high-volume sarcoma centers remains essential for optimal patient care. At the same time, LLMs could serve as an excellent tool to prepare for MDT discussions.