Researchers from UHN’s Ajmera Transplant Centre have simulated an organ transplant selection committee using an AI framework. The study, published in The Lancet Digital Health, found that an AI system made up of several specialized components, called “agents”, could accurately identify which patients were likely to benefit from a liver transplant and which were not.

Liver transplantation is offered as a life-saving intervention for patients with end-stage liver disease. However, the demand for organs often exceeds the supply available for transplantation. Candidacy for a transplant is considered based on need and survival benefit for patients.  

Decisions about which patients should be placed on liver transplant waiting lists are currently made by multidisciplinary transplant selection committees. Despite established guidelines to ensure fairness, equitable access, and standardization, as well as the best intentions of clinicians involved, subjectivity and bias can still impact the process. This can result in variability in decisions within and between institutions, especially for complex cases. Therefore, there is a need for greater objectivity and consistent, data-driven decision-making.

Large language models (LLMs)—AI systems trained on large amounts of text that can read, interpret, and generate human-like language—are increasingly being used as tools in health care. LLMs can act as AI agents—systems capable of autonomously performing tasks on behalf of a user. They can also be used in multi-agent collaborations, where multiple AI agents work collectively to help solve complex problems.

Although multi-agent frameworks have the potential to support health care decision-making for clinical teams, their application in the medical field remains limited.

To investigate the use of multi-agent AI as a way to aid in objective health care decision-making, the researchers examined the performance of a multi-agent AI selection committee (AI-SC) in the liver transplant selection process. This AI committee was designed to reflect a typical transplant team, with each AI agent assigned the role of one member of the team, including a transplant hepatologist, transplant surgeon, cardiologist, or social worker.

The AI-SC then assessed whether patients should be added to the transplant waiting list by determining whether a liver transplant was likely to improve survival at six months or one year. Patients were declined if medical factors made transplantation inappropriate to consider or unlikely to extend life.

The AI-SC was tested using information from a database of approximately 8,400 liver transplant candidates from 2004 to 2023. Researchers analyzed the outcomes of patients who received a transplant, and then created a simulated group of patients, including patients with medical conditions that would automatically rule out transplantation, for the AI-SC to assess.

Results showed that the AI-SC triaged potential transplant patients with high accuracy. It correctly identified patients who should not receive a transplant 98% of the time, and predicted whether a transplant would provide a survival benefit at both six months and one year after surgery with accuracies of 92% and 95%, respectively.

There were still instances where the AI-SC declined patients who had ultimately been accepted (false negatives) or accepted patients who died before 6 months or a year post-transplant (false positives). These false negatives and false positives were largely related to underlying cancer‑related factors and greater complexity in the patients' medical history.

Overall, the findings suggest that multi-agentic AI could potentially support transplant teams by helping them make consistent, evidence‑based decisions—improving the transplant selection process. This UHN-developed tool is the first instance of a multi-agentic AI system to guide objective, consensus-based decision-making in health care. As an important next step, the Transplant AI Initiative team is currently evaluating this tool at an international, multicenter level to assess its role across different clinical contexts.

Dr. Bima Hasjim is an International Clinical Fellow in the Bhat lab and a surgical resident at the University of California. He is a co-first author of the study.

Dr. Ghazal Azarfar is a Research Associate in the Bhat lab and co-first author of the study.

Dr. Divya Sharma is an AI Scientist and Clinician Investigator at UHN’s Princess Margaret Cancer Centre. She is an Assistant Professor at York University in the Department of Mathematics and Statistics, with an additional appointment at the Dalla Lana School of Public Health at the University of Toronto. She is a co-senior author of the study.

Dr. Mamatha Bhat, Scientist at UHN’s Ajmera Transplant Center and Associate Professor in the Department of Medicine at the University of Toronto, is a co-senior author of this study.

This work was supported by the Transplant AI Initiative, Ajmera Transplant Centre, and UHN Foundation.

Dr. Bhat has received grants/contracts from Novo Nordisk, Paladin, Oncoustics, Merck, and Transplant Genomics, payment/honoraria from Paladin, and reports participating on a Data Safety Monitoring Board/Advisory Board for Novo Nordisk.

Hasjim BJ, Azarfar G, Lee FG, Diwan TS, Raju S, Gross JA, Sidhu A, Ichii H, Krishnan RG, Mamdani M, Sharma D, Bhat M. A multiagent large language model–based system to simulate the liver transplant selection committee: a retrospective cohort study. Lancet Digit Health. 2026 Apr 7;100966. doi:10.1016/j.landig.2025.100966. Epub 2026 Apr 7.