Interpretable rules guide kidney allocation decisions
When Should AI Decide, Defer, and Deviate? Responsible Automation for Deceased Donor Kidney Allocation · arXiv:2610.39045
In plain words
U.S. kidney offers usually follow a priority list. Moving a kidney out of order may help place one that would otherwise go unused, but it can bypass patients higher on the list.
The Arbitrator is a decision framework that predicts whether a kidney would be finally accepted at original ranked match-sequence position 100 or earlier, then uses interpretable rules to recommend keeping it in sequence or considering a move.
In retrospective tests, its rules resolved 93.5% of 21,574 kidney decisions; the rest were deferred. That measures agreement with historical acceptance patterns, not transplant success.
The research
A kidney offered in priority order may go unused while a match is sought, yet moving it out of sequence can bypass higher-priority patients. Can a decision aid identify kidneys that may need faster placement while preserving the standard order when it is likely to work?
The Arbitrator certifies stable, interpretable routing rules, defers uncertain cases, and lets a literature audit flag rules before possible automation.
The Arbitrator uses interpretable models to recommend whether a kidney should stay in the standard offer sequence or be considered for out-of-sequence placement. It certifies an action only when the models’ uncertainty does not change which action is preferred, and defers unresolved cases.
On 21,574 held-out test kidneys, rules resolved 93.5% of decisions. Average decision cost was 0.143, compared with 0.131 when XGBoost decided every case.
This supports the framework’s ability to provide interpretable decisions for most test cases, while showing a higher cost against historical acceptance labels.
This was a retrospective study: acceptance at original match-sequence position 100 or earlier is a proxy for placement difficulty, not transplant success, and the framework does not choose a kidney’s destination. The literature audit was exploratory and needs clinical validation.
What the evidence shows
- Balancing placement and priority
The Arbitrator is a decision framework for deceased-donor kidney allocation. Standard offers follow a priority sequence, but a difficult-to-place kidney may go unused while waiting; moving it out of sequence may help, while bypassing higher-priority patients raises fairness concerns.
The framework aims to identify when to keep the standard sequence and when to consider an out-of-sequence route, while deferring uncertain cases.
- Define the placement signal
To estimate placement difficulty, the authors predict whether a kidney would be finally accepted at original ranked match-sequence position 100 or earlier. They call this early-matchability.
It is a proxy based on historical acceptance timing, not a prediction of transplant success. The framework considers out-of-sequence placement when this predicted probability is low, but it only certifies a routing action when its uncertainty interval cannot change which action is preferred.
- Certify or defer
Two interpretable models contribute rules: a decision tree and a compact set of rules drawn from a random forest. When both support the same action, or one supports an action while the other abstains, the Arbitrator resolves the decision; it defers when neither action is certified or their certifications conflict.
In the held-out test set of 21,574 kidneys, rules resolved 93.5% of decisions at the study’s chosen cost setting.
- Read coverage with its cost
That coverage came with a higher average decision cost than using XGBoost for every test kidney: 0.143 versus 0.131. The comparison uses historical acceptance labels; XGBoost represents deferred cases in the Arbitrator evaluation because actual expert review was not observed.
The framework also recommended out-of-sequence placement for 1,920 of 2,738 historically discarded low-KDPI test kidneys, a retrospective recommendation rather than a count of successful transplants.
- Add a literature review
A separate audit asked a GPT-6-based OpenAI Codex agent to compare 113 certified rules with selected medical literature and guidance. It recommended retaining 62 actions provisionally, referring 43 rules for physician review, and reconsidering the opposite action for eight.
Applying these recommendations through a hypothetical procedure left 86.6% of originally resolved test kidneys eligible for automation, conditional on the recommendations being well-founded.
Deep
- Original paper authors
- Madeleine Pollack, Thomas Daillak, Mehmet Toner, Korkut Uygun, Heidi Yeh
- Submitted to arXiv
- License
- CC BY 4.0