KVIG Informative

A donated organ starts with a hospital recognizing that a dying patient could be a donor and calling the right people in time. Software is now being built to do that step faster, and researchers are testing whether machine learning can help.

The Referral Step When a patient is near death, the hospital is supposed to notify its organ procurement organization (OPO) so trained staff can assess whether donation is possible and approach the family. Historically that call depended on a nurse or physician remembering to make it. UNOS, the contractor that has run the OPTN, is asking Congress to require automated electronic referrals, in which the hospital’s electronic health record flags a patient showing signs of imminent death and sends the OPO a message with enough information to begin evaluating suitability.

Automation Is Not the Same as AI UNOS says the technology already exists and pilot programs are underway. Its position paper describes software and automation, and it does not use the term AI. This distinction matters: a rule that says “alert the OPO when a ventilated patient meets these criteria” is a fixed trigger. A machine-learning model instead learns patterns from past cases and estimates a probability. Both are in play, and they carry different risks.

A Real Example Ohio State University Wexner Medical Center announced in July 2023 that it had launched a digital donor referral tool called iReferral, from the company InVita, at 15 of its facilities, working with Lifeline of Ohio, the OPO for central and southeast Ohio. The announcement did not describe the tool as using AI. InVita’s president and CEO, Todd Collins, said the tool streamlines the referral process.

What Machine Learning Has Shown A research paper posted on the medRxiv preprint server used the ORCHID dataset from PhysioNet: 133,101 donor referral records from six U.S. OPOs between 2015 and 2021. The authors tested seven classifiers on two questions, whether a referred donor would be selected for transplant and whether a selected donor would yield multiple organs. The best model for the first question, XGBoost, reached a recall of 0.80 and an F1 score of 0.83. The second question was easier: several models scored at least 0.95 on recall, precision and AUC. The strongest predictors for selection were administrative milestones, especially approaching the donor’s family.

The Limits The same paper lists its limits. The dataset lacks comprehensive donor and administrative information, much of the missing data is not random, and timestamps were randomly shifted for privacy. The authors also flag that recall was lower than precision for the first question, which raises the risk of false negatives, meaning a model could pass over someone who would have donated. It is a preprint, so it has not completed peer review. A model that mostly learns that families who are approached usually say yes would not tell an OPO anything it does not know.

The Nevada Picture Nevada Donor Network, the OPO that serves Nevada, reported 795 organ transplants and 262 organ donors across the state in 2025. Its president and CEO, Joseph Ferreira, credited hospital and community partners. CMS describes the Las Vegas-based OPO as Tier 1 rated, and on Jan. 8, 2026, CMS approved it to take over organ procurement in the Southern Florida service area, which covers about 7 million people. The sources reviewed do not say whether Nevada Donor Network uses machine-learning tools.

Readers who want to know how their own hospital or OPO handles referrals can ask them directly. The useful questions are simple: is the referral automatic, and who checks that it happened?