Multidisciplinary large language model agent teams for precision oncology enhance complex gynecologic oncology decision support
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Large language models can help with clinical decision-making tasks. Complex oncology cases are best managed through multidisciplinary tumor boards but are difficult to do so due to their expense. The MDAT framework is proposed to mimic tumor board-style collaboration. Some LLMs are prompted to act like experts. They first analyze the prompt from their respective perspectives. Then the decision-making takes place to vote, agree, deal with discord and unify. We evaluated this framework on 3 LLMs: ChatGPT-4o, DeepSeek-R1 and Llama-4 (1,056 clinical questions; 182 cases). It focused on staging and selecting the treatment, managing the complications and following up. Fixed-size MDATs, across all baselines and configurations of MDAT, outperformed six different strong prompting baseline algorithms across all models. DeepSeek-R1 benefits from a five-agent MDAT that performs best (98.26/100). The fixed-size MDATs showed the greatest gains in higher-complexity questions, demonstrating robustness where accurate, multidisciplinary reasoning is most needed. As our MDT-inspired agent enhances LLM accuracy for oncology decision-making, it provides a pragmatic approach to integrating AI into complex cancer management.
Author summary
Experts, such as surgeon, oncologist and radiologist get together to decide on the best treatment for a cancer patient but it is rather costly and time consuming. Large language models exhibit excellent performance when it comes to answering questions, particularly in medical contexts. We wanted to discover if large language models can work collaboratively in the same manner. We designed a framework named MDAT, standing for Multidisciplinary Agent Team. Rather than asking a question to one generic AI, our system builds a team of AIs “experts”. Each agent considers a patient’s case from its perspective before all of them work together in a well-designed system to provide the expert recommendation. Through multidisciplinary team (MDT) gynecologic cancer scenarios, we examined this team-based artificial intelligence on a custom-built dataset of more than 1000 questions. According to our findings, this method always provides more reliable and accurate answers than decision-making methods, especially for the toughest cases. According to our findings, we can create safer and more effective ways of helping doctors make crucial decisions for cancer patients by making AI act like a real medical team.
