KVIG Informative

Once an organ is offered, a surgeon has to decide whether it is good enough to use. Researchers are building tools to help with that call, to predict which offers will be accepted, and to test allocation rules before they take effect.

Checking a Liver From a Photo Surgeons often judge how fatty a donor liver is by eye, and fat content affects whether the organ is likely to work. A research team developed a platform called LiverColor that extracts color and texture features from photographs of donor livers and uses machine learning to classify each graft as valid or not valid. According to the study, published in the journal Diagnostics in 2024, the models were built on 192 cases, compared against biopsy results and surgeons’ classifications, and reached an AUC of 0.82 and 85% accuracy. The authors report that this exceeded surgeons’ visual inspection, and that experienced surgeons rated it positively in simulated and real usability testing.

What This Does and Does Not Prove The result comes from one in-house dataset of 192 cases. A tool that works well in one hospital’s photos, lighting and camera can perform differently elsewhere, and the study as summarized does not show that it improves patient outcomes. It is a decision aid for a surgeon, not a replacement for one.

Predicting Who Says Yes Many organs are offered to several patients before one is accepted. A team at Microsoft Research, including Carlos Martinez, Juan M. Lavista Ferres and Bill Weeks, published work in the Journal of Transplantation on predicting whether a kidney offer would be accepted. They compared logistic regression with gradient-boosted trees and found XGBoost worked best, with an average precision of 0.0907. Adding nonclinical data, such as the distance from the nearest airport to the transplant hospital, raised it slightly to 0.0940. The baseline AUROC of 0.832 looked strong, but the low average precision showed the models struggled because declines far outnumber acceptances.

The Authors’ Own Warning The paper’s bottom line is cautious. Even the best models did not clearly or equitably outperform existing allocation policies, the authors write, and the data contained nothing about the human reasoning behind each offer decision. They conclude that more research is needed before clinical use. That is a useful standard for any tool that claims to improve matching.

Testing Rules Before They Start The MIT Transplants AI project and Dynamic Ideas have built a simulator of the U.S. transplant system. It starts from a snapshot of a historical waiting list, applies an allocation policy, and simulates offers, acceptances and outcomes, so analysts can compare waitlist death rates and access across groups. The team also uses multi-objective optimization to search for weights that best meet policymakers’ goals. The project page does not say the simulator has been used in OPTN decisions.

Taking Inventory On April 30, 2026, the Organ Donation and Transplantation Alliance launched an AI Resource Hub that maps AI work across donation and transplantation and asks innovators to submit pilots, findings and lessons. Its announcement lists no specific tools and states no cautions, so the hub is a directory to watch rather than evidence that any tool works.

The pattern across these projects is the same: promising results on historical data and a gap before use on patients. The next article looks at the part of that gap where harm can occur, which is bias and who is responsible for checking.