Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
AI-Powered Liver Transplant Selection: Multiagent Systems Show Promise for Objective Allocation
An AI committee of specialized agents using GPT-4o accurately simulated liver transplant selection decisions, potentially reducing bias and inconsistency in this resource-limited allocation.
Background
Liver transplantation requires committees to select candidates who will benefit most while identifying those with absolute contraindications. These complex multidisciplinary decisions can be subject to bias and inconsistency across centers.
Key Findings
- AI committee achieved 98.19% accuracy identifying contraindications (100% sensitivity, 91% specificity)
- 1-year survival prediction: 92% accuracy (100% sensitivity, 66% specificity)
- 6-month survival: 94.88% accuracy (100% sensitivity, 75% specificity)
- Of 8,412 clinical vignettes from real SRTR data, 83.6% were recommended for transplant listing
- Fairness maintained across demographic groups (disparate impact scores ≥0.960)
Why It Matters
This multiagent framework could reduce bias and variability in organ allocation while supporting objective, multidisciplinary reasoning in high-stakes clinical decisions. The approach demonstrates AI’s potential to handle complex resource allocation where consistency and fairness are paramount.
Limitations
The study used simulated clinical vignettes rather than direct comparison with real committees. Human oversight remains essential, and explainability of AI reasoning and accountability mechanisms require careful consideration before clinical implementation.
Original paper: A multiagent large language model-based system to simulate the liver transplant selection committee: a retrospective cohort study. — The Lancet. Digital health. 10.1016/j.landig.2025.100966
