Liver transplant donor-recipient matching with offline reinforcement learning

Area of Science:

• Medical Informatics
• Machine Learning in Medicine
• Transplantation Surgery

Background:

• Liver transplantation (LT) faces challenges due to high waitlist mortality, organ scarcity, and graft failure risk.
• Current methods statically predict successful donor-recipient pairs, neglecting dynamic risks over time.
• Optimizing LT decisions requires balancing waitlist mortality against graft failure risks.

Purpose of the Study:

• To apply an offline reinforcement learning (RL) approach for optimizing sequential decisions in liver transplantation.
• To dynamically manage patient waitlist status (wait, delist, transplant) based on evolving risks.
• To improve donor-recipient matching and patient outcomes in LT.

Main Methods:

• Utilized an offline reinforcement learning (RL) framework to model LT decision-making.
• Trained the RL model using historical waitlist data from the Scientific Registry of Transplant Recipients (SRTR) database.
• Represented the LT process as a series of time-dependent decisions: wait, delist, or transplant.

Main Results:

• The RL model avoided 73% of donor-recipient pairs that would have resulted in graft failure or death.
• Successfully preserved 93% of transplants that led to positive outcomes.
• Identified potentially suitable donors for 47% of patients who died while on the waitlist.

Conclusions:

• RL-based approaches offer a more dynamic and realistic model for liver transplantation donor-recipient matching.
• The model learned key features indicative of successful donor-recipient pairs, enhancing clinical decision support.
• This study demonstrates the potential of RL as a valuable clinical tool for improving LT outcomes.