A new international study involving 첥Ƶ researcher expertise shows that AI could help make liver transplant decisions more consistent, transparent and evidence-based, especially when resources are limited.
The study, published in , tested a multi-agent system built with large language models (LLMs) to simulate the work of a liver transplant selection committee – a multidisciplinary group that decides which patients are placed on transplant waitlists.
Using real-world transplant registry data, the AI system demonstrated high accuracy in identifying patients who are likely to benefit from a liver transplant and those for whom transplantation would be unlikely to help.
“Liver transplantation is a rare case in medicine where access to a life-saving treatment is limited by organ availability,” explains co-senior author Divya Sharma, assistant professor in the Faculty of Science. “Decisions about who is waitlisted are complex, and committee deliberations can be subject to unconscious bias where a clinician's own background or identity may subtly influence their judgement, even when national guidelines are in place.”
Researchers set out to test whether AI agents – each assigned a clinical role – could support more objective decision-making. To test the approach at scale, researchers evaluated the system against transplant outcomes data.
The study analyzed 20 years of data from more than 8,000 adult liver transplant recipients in the U.S. using the Scientific Registry of Transplant Recipients. A simulated group of patients with known contraindications was also created to test the system’s accuracy in flagging cases that should be excluded from transplant consideration.
Results show the AI committee predicted one-year post-transplant survival with 92 per cent accuracy and six-month survival with 95 per cent accuracy. Contraindications were identified with an accuracy of more than 98 per cent, thereby identifying transplant candidates efficiently.
The research team also examined where errors occurred to better understand where the AI system works well, and where it needs careful oversight and improvement. The authors caution that continued monitoring is needed because transplant data can reflect broader inequities in access to health care.
“Our work positions LLM-based multi-agent AI systems as potential clinical decision-support tools, rather than replacements for human judgement,” says Sharma. “While AI shows promise in making liver transplant decisions more objective, it’s crucial to emphasize that the final responsibility must always remain with transplant teams and human oversight is critical to address ethical considerations.”
Sharma says while more research is needed to test the AI tools in real-world settings across different health systems, AI-supported committees have potential to help standardize complex medical decisions where resources are limited.