AI in Organ Transplantation: Matching, Allocation, and Saving More Lives

In 2023, American surgeons performed a record 28,142 kidney transplants, and in the same year 27.9% of the deceased-donor kidneys that were recovered for transplant were never used. Both numbers sit inside the same system. Machine learning, the branch of AI that finds patterns in past records, is now being tested on the decisions that sit between a donor and a recipient: who gets an offer, which organ is worth accepting, and which patient is most likely to die waiting. This article explains what those models do, what the evidence really shows, and why fairness is the hard part.

Educational content, not medical advice. Talk to your doctor about any decision involving a transplant, a medication, or a treatment plan. This article describes research and policy, and every patient's situation differs.

Key takeaways

• Published machine learning models predict some transplant outcomes, such as waitlist mortality, with higher accuracy than established scores in retrospective tests, but the evidence is mostly simulated or retrospective rather than from prospective use.
• Kidney nonuse is a decision problem as much as a supply problem: in 2023, 72.5% of kidneys with a KDPI (Kidney Donor Profile Index) of 85% or higher went unused.
• Simulation studies of liver allocation report fewer waitlist deaths under machine learning prioritization than under the MELD score, but simulation is not a clinical trial.
• Every allocation rule encodes values. AI makes the trade-offs between urgency, benefit, and fairness easier to measure, and does not decide which trade-off is right.

Why a donated kidney can still go unused

The word "discard" hides a long chain of decisions. A kidney is recovered after a donor dies, offered to candidates in a ranked order, and declined by one transplant center after another until it is either accepted or abandoned. The official term for the abandoned outcome is nonuse, meaning the organ was recovered for transplant and then not transplanted. The scale of it is the surprise: according to the OPTN/SRTR 2023 Annual Data Report on kidneys, the share of recovered deceased-donor kidneys that went unused rose to 27.9% in 2023, up from 26.6% in 2022 (Europe PMC record of the report).

Nonuse is not spread evenly. The same report found it was far higher for kidneys from donors aged 65 or older, for kidneys that had been biopsied, and for kidneys with high KDPI scores. A kidney does not have to be bad to land in these categories; some of them function well for years in the right recipient. The decision to decline often reflects risk-aversion, logistics, and the fact that each center sees only its own patients.

Kidney category (US, 2023)	Share of recovered kidneys not used
All recovered deceased-donor kidneys	27.9%
Same group in 2022	26.6%
Biopsied kidneys	41.4%
Donors aged 65 or older	72.2%
KDPI of 85% or higher	72.5%
Hepatitis C NAT-positive donors (2023)	27.2%
Hepatitis C NAT-positive donors (2017)	43.0%

Source: OPTN/SRTR 2023 Annual Data Report: Kidney, American Journal of Transplantation, published 2025.

The last two rows are the most hopeful. Nonuse of kidneys from donors with hepatitis C fell from 43.0% in 2017 to 27.2% in 2023, which the report's authors link to wider use of direct-acting antiviral drugs that can cure the infection after transplant. That drop came from better medicine and changed protocols, not from an algorithm. It shows what a better decision framework can do when the risk is understood.

Think of an airline with an empty seat on a departing flight. The seat is worth something until the door closes, then it is worth nothing. A kidney works the same way, because the organ loses viability while it waits, so the cost of a slow or overly cautious decision is paid by the patient who would have received it.

How allocation works today, and where rules strain

In the United States, organ allocation is run by the Organ Procurement and Transplantation Network (OPTN), a federally contracted system that has long been operated by the United Network for Organ Sharing (UNOS). When an organ becomes available, the system generates a ranked list of candidates using rules written separately for each organ. Those rules weigh blood type compatibility, the immune match between donor and recipient, geography, time on the waiting list, and medical urgency.

Each factor has a biological reason behind it. Blood type matters because the immune system attacks cells that carry the wrong sugar markers, the same reason transfusions are matched (blood types and the ABO system). The immune match matters because human leukocyte antigen (HLA) genes build the proteins that display fragments on cell surfaces, and T cells use those displays to tell self from foreign (how T-cell receptors read those displays). Geography matters because an organ without blood flow is a clock: hearts and lungs tolerate far fewer hours than kidneys do.

The strain comes from how many objectives pull in different directions at once. A rule can favor the sickest patient, the patient who will gain the most years, the patient who has waited longest, or the patient who is closest to the donor hospital. Rules combine these with weights, and the weights are negotiated by committees and revised over years.

Allocation is not a single decision either, which makes it harder to improve. There are three linked choices:

1. Whether the donor organ should be accepted for transplant at all.
2. Which candidate on the list should receive an offer first.
3. Whether that candidate's team should say yes, given the patient's health today and the chance of a better offer later.

A model that improves only the second choice can still fail if the third choice is where organs are lost. This is why predictive accuracy, taken alone, is a weak guide to whether an AI tool will save lives.

Predicting graft survival: what scores do and what machine learning adds

Two scores shape kidney decisions. The Kidney Donor Profile Index (KDPI) summarizes a deceased donor's characteristics into a percentile where a higher number means shorter expected graft life. The Estimated Post-Transplant Survival (EPTS) score does something similar for recipients. Both were built with classical regression, which fits a small number of variables to outcomes using relationships that are assumed to be smooth and mostly additive.

Machine learning methods, such as gradient boosting and neural networks, can pick up interactions that regression misses. Gradient boosting builds hundreds of small decision trees in sequence, each one correcting the mistakes of the previous ones (how gradient-boosted trees such as XGBoost work). The usual way to compare models is the area under the curve (AUC). An AUC of 0.5 is a coin flip and 1.0 is perfect ranking, so a model that moves from 0.65 to 0.75 is ordering patients by risk noticeably better.

The honest summary of the field is more cautious than the headlines. A 2025 systematic review in BMC Medical Informatics and Decision Making examined machine learning for liver allocation across studies published from 2002 to June 2023. It included 20 papers, and eight of them came from one Spanish group working on donor-recipient matching and post-transplant survival. The authors found that most studies focused on predicting utility, and that robust models of transplant benefit, and urgency models that outperform MELD, were lacking (Pruinelli et al., 2025).

That finding is the standard pattern in clinical AI. A model can score well on the data it was built from and lose accuracy at a new hospital, in a new decade, or under a new policy. One common cause is data leakage, where information from the future quietly enters the training data and inflates the results (why data leakage breaks AI evaluations). A transplant registry is especially exposed because the same rules that created the data also shaped who got which organ.

How to read a claim that a model beats the current score

Headlines about transplant AI usually compress four separate questions into one, and each deserves its own answer. Knowing the questions makes it easier to tell a strong study from a weak one.

Was it tested on patients the model never saw? A model is fitted on a training set, then checked on a held-out set. The strongest design adds a third check on a different country or health system. The GEMA-AI liver model followed that pattern: 5,762 candidates for training, 1,920 for internal validation in the UK, and 1,638 for external validation in Australia. A model that holds up in Australia after being built on other data is less likely to have memorized quirks of one registry (why training and test data must be separated).

Does it rank well, and does it estimate well? Discrimination is the ability to put higher-risk patients above lower-risk ones, which AUC measures. Calibration is a separate property: when a model says 20% risk, do about 20 in 100 such patients actually die? A model can rank patients perfectly and still give probabilities that are too high or too low, which matters when a threshold decides who gets an offer.

Who gains, and who loses? An average improvement can hide a group that does worse. The GEMA-AI study reported its gain by sex, finding the larger benefit among women. Every allocation model should be reported this way, by sex, by race, by diagnosis, and by region, because a rule that helps the average candidate can still move organs away from a particular group.

What is the outcome, and is it the right one? Predicting death within 90 days is not the same as predicting years of life gained from a transplant. A candidate who is very sick and likely to die soon may also be a patient for whom a transplant carries a high early risk. Urgency and benefit are different quantities, and the 2025 systematic review found that robust models of transplant benefit were the missing piece.

A useful comparison is a restaurant critic who visits only on Friday night. The reviews may be accurate for Friday night and misleading about the Tuesday lunch crowd. A transplant model trained on one era, one set of rules, and one population has the same blind spot, and the fix is the same: check it somewhere else.

Liver allocation: where the simulation evidence is clearest

The liver offers the cleanest example of machine learning aimed at allocation, partly because the existing score is so well known. The Model for End-Stage Liver Disease (MELD) has ranked liver candidates by short-term mortality risk for more than 20 years. It uses a handful of blood tests, and critics have long noted that it can under-prioritize groups, including women (the liver's role in the body).

In 2019, a team led by Dimitris Bertsimas at MIT published the Optimized Prediction of Mortality (OPOM) in the American Journal of Transplantation. OPOM is a machine learning classification tree that predicts a candidate's chance of dying or being removed from the list within three months. The researchers then ran OPOM through the Liver Simulated Allocation Model, a computer simulation of the national waiting list. They estimated that allocating by OPOM instead of MELD would prevent about 418 deaths per year, with a reported range of 406.8 to 428.4, and that survival would improve across demographic groups, diagnoses, and regions (Bertsimas et al., 2019).

A 2025 study in Clinical Gastroenterology and Hepatology took a related approach with a different goal. GEMA-AI is an explainable neural network built to correct the score's disadvantage for women. It used 9,320 adult waitlist candidates, split into 5,762 for training, 1,920 for internal validation in the UK, and 1,638 for external validation in Australia. It outperformed GEMA-Na, MELD 3.0, and MELD-Na, with a larger benefit for women, and the authors estimated it could prevent 1 in 59 deaths overall and 1 in 13 deaths among women (Gomez-Orellana et al., 2025).

Both results are encouraging, and both are simulations or retrospective analyses. A simulated death prevented is a model output. It depends on assumptions about how centers would behave under a new rule, and real centers adapt to rules in ways simulations can miss.

Reducing nonuse: biopsies, marginal donors, and machine perfusion

Kidneys from older donors, or from donors with hypertension or diabetes, were once grouped as "expanded criteria" organs. The label is a population average. It says that kidneys like this tend to have shorter lives, and says little about whether this kidney will serve this recipient well. For an older patient who may spend years on dialysis, a shorter-lived kidney offered today can be a better bet than a better kidney that arrives in two years, if it arrives at all.

This is the logic behind the "benefit over alternatives" framing. A decision model compares a recipient's expected outcome with this organ against the expected outcome of waiting. Prediction models that estimate both sides are an active research area. Published results for them are mostly retrospective, so the numbers in this article come from registry reports, trials, and peer-reviewed simulations rather than from product claims.

Hardware may change the decision more than software. Normothermic machine perfusion keeps an organ warm and supplied with oxygenated blood outside the body, in place of the standard method of packing it in ice. A randomized trial of 220 liver transplants published in Nature in 2018 found that, compared with cold storage, normothermic preservation was associated with 50% lower graft injury as measured by enzyme release, a 50% lower organ discard rate, and a 54% longer mean preservation time. There were no significant differences in bile duct complications, graft survival, or patient survival (Nature, 2018).

Perfusion matters to AI because it produces data. While the organ is on the machine, sensors record flow, temperature, and chemistry, and the fluid can be tested for markers of cell injury. A model trained on those streams could, in principle, tell a marginal-looking organ that will recover from one that will not. That is a plausible research direction. What has been tested in a randomized trial is the perfusion hardware itself, not a validated prediction model built on its sensor data.

Simulation and optimization: designing the rules, not just the predictions

Predicting one patient's outcome is a different job from choosing a policy. A policy decides, across thousands of patients and organs, which offers to make and in what order. Policy design resembles air traffic control more than weather forecasting. A forecast describes what will probably happen. Air traffic control decides what should happen, knowing that every decision changes what the next one looks like.

The tool for this job is simulation. The OPTN uses simulators that replay historical donors and candidates under proposed rules and count the results. They are slow. A 2025 paper in Transplantation Direct by Pivo, Bertsimas, Trichakis, and colleagues describes a new algorithm that cut the time to simulate one year of kidney allocation from more than six hours to about 15 seconds, using the same underlying model as the existing OPTN simulator. That speed allowed the team to test thousands of policies with multiobjective optimization, run 50,000 simulations, and highlight four candidate policies that balance competing goals in different ways (Pivo et al., 2025).

The work supports an effort called continuous distribution, in which candidates receive a composite score built from several attributes rather than being sorted into rigid tiers. Reinforcement learning, the technique in which software learns a strategy by trial and error against a simulated environment, is a natural fit for sequential offer decisions (how reinforcement learning works). The published work on kidney policy describes simulation and multiobjective optimization, not a reinforcement-learning agent making live offers. A reader who sees a claim that an RL agent "saves thousands of life-years" should ask whether that is a simulated result and what its assumptions were.

Predicting rejection: HLA matching, antibodies, and cell-free DNA

Even a well-matched organ can be attacked. Rejection comes in two broad forms. T cell-mediated rejection occurs when immune cells recognize the graft as foreign and attack it. Antibody-mediated rejection occurs when antibodies bind the organ's blood vessels and trigger damage (how antibodies are built and what they do). They are treated differently, which is why distinguishing them matters.

The standard tool for diagnosis is a biopsy, which is invasive and can miss patchy damage. A less invasive approach measures donor-derived cell-free DNA (dd-cfDNA). When cells in the transplanted organ die, they shed fragments of DNA into the recipient's blood, and the fraction that comes from the donor rises during injury. The idea resembles a liquid biopsy in cancer care (liquid biopsy explained).

The foundational study, published in the Journal of the American Society of Nephrology in 2017 by Bloom and colleagues, included 102 kidney recipients and 107 biopsy specimens. At a cutoff of 1.0% dd-cfDNA, the test detected active rejection with a positive predictive value of 61% and a negative predictive value of 84%. For antibody-mediated rejection, the positive predictive value was 44% and the negative predictive value 96%, and the area under the curve was 0.87 (Bloom et al., 2017).

Those figures show how to read such a test. A negative result is reassuring, because 96% of patients with a low value did not have antibody-mediated rejection. A positive result is only a prompt to look closer, because more than half of the positives for that form of rejection were false alarms. Median values told a similar story: 2.9% in antibody-mediated rejection, 1.2% in higher-grade T cell-mediated rejection, 0.2% in lower-grade cellular rejection, and 0.3% in controls. Commercial assays from companies such as CareDx use this principle, and machine learning enters mainly in the interpretation layer, where it combines the fraction with clinical history, antibody profiles, and trends over time.

Where bias enters: KDPI, race, and sex

Any model learns from records, and records carry history. A candidate who had less access to specialists, to transplant centers, or to medication support may have worse recorded outcomes, and a model trained on those outcomes can treat the pattern as a fact about biology. The risk is a feedback loop. A score ranks a group lower, the group gets fewer good organs, its outcomes look worse, and the next version of the score ranks it lower still (how algorithms fail through biased data).

A concrete case came from the kidney donor score itself. The Kidney Donor Risk Index, which feeds KDPI, included a variable for whether the donor was Black. A 2023 analysis in the American Journal of Transplantation recalculated the index without that variable. It used 69,244 deceased-donor kidney transplants from 1995 to 2005 to refit the coefficients, and applied them to 72,926 deceased donors from 2015 to 2021. Removing race had a negligible effect on the model's ability to predict outcomes, but it changed the ranking: the share of Black donors with a KDPI above 85% fell from 31.09% to 17.75%, close to the 15.68% for non-Black donors (2023 analysis in the American Journal of Transplantation).

Combine that with the nonuse table above. Kidneys with a KDPI of 85% or higher were left unused 72.5% of the time in 2023. A score that pushes more kidneys from one group into that band raises the chance those kidneys are never used. The arithmetic suggests that the race term did not make predictions meaningfully better, and may have cost organs. Fixing it required a human decision to remove a variable, not a more powerful algorithm.

Sex bias is the liver counterpart. The 2026 review in Liver Transplantation by Reddy and colleagues notes that disparities by sex, race, and structural factors persist under MELD, and it argues that AI should support rather than replace clinical decisions. It proposes a governance framework covering fairness monitoring, transparency, explainability, human oversight, and multi-stakeholder governance (Reddy et al., 2026).

A hypothetical offer, step by step

The following example is invented for illustration and uses no real patient data. It shows where a model could and could not help.

A kidney becomes available from a 67-year-old donor with long-standing hypertension. Its KDPI falls in the high band, the range where 72.5% of kidneys went unused in 2023. The ranked list contains a 34-year-old whose dialysis access is failing, a 62-year-old with diabetes, and a 71-year-old who has already waited several years.

Under today's rules the offer goes first to the highest-ranked candidate. A transplant team will often decline a high-KDPI kidney for a young recipient, reasoning that the patient may need a long-lived organ and can wait for a better one. That reasoning is sensible, but it assumes a better offer will actually arrive. If the young patient's antibody profile makes compatible donors rare, the wait could be long, and the decline may cost more than the risk it avoids.

A decision-support model could estimate, for each candidate, the expected outcome with this kidney and the expected outcome of waiting, using the candidate's own data. The 71-year-old may gain the most from accepting now, because a shorter-lived kidney still outlasts the expected remaining years. The model's role is to produce those two estimates and show the uncertainty around them. It does not decide.

A human still has to ask the questions a model cannot answer. Does this candidate want this kidney? Is the center able to manage the added risk? Does the ranking rule, which gives weight to the young patient's longer expected life, reflect a value the public supports? Those are the points where policy committees, not software, set the terms.

The example also shows why timing matters. Every hour spent on declines is an hour of cold storage, and the longer an organ waits the worse its prospects become. A tool that returned estimates within minutes could help more than one that was slightly more accurate and slow.

What stops AI from running allocation today

Several barriers keep these models in research. They are worth listing because each one is a practical reason, not a technicality.

• Evidence type. Most results come from retrospective data or simulation. A prospective test of a new allocation rule is slow, risky, and hard to randomize.
• Interpretability. Committees, patients, and courts need to understand why one person ranked above another. A black-box score is a hard sell in a system where the ranking is a matter of life and death.
• Data quality and drift. Registry data are entered by many centers and shift as practice and policy change. A model trained on 2015 data may not describe 2026.
• Gaming and behavior. When a rule changes, clinicians change what they record and whom they list. A model that predicts well under one rule can mislead under another.
• Governance. Allocation policy is public policy. The weights between urgency, benefit, and fairness are value choices that a committee must own and defend.

A 2026 narrative review of AI in lung transplantation, published in the Journal of Thoracic Disease, reached similar conclusions across a different organ. It found that AI could shift allocation from urgency toward utility and help with donor assessment and complication prediction, that post-transplant applications were the most mature, and that data quality, interpretability, and integration into clinical practice were the main barriers (Luo et al., 2026).

What each AI approach does, and how strong the evidence is

Use case	What the model does	Strongest published evidence	Main limit
Waitlist mortality (liver)	Ranks candidates by 3-month death risk	OPOM simulation: about 418 deaths per year averted vs MELD (2019); GEMA-AI outperformed MELD 3.0 in 9,320 candidates (2025)	Simulated benefit; real-world behavior may differ
Allocation policy design	Tests thousands of rule variants quickly	Kidney simulator cut from 6+ hours to about 15 seconds per year; 50,000 simulations (2025)	Depends on the simulator's assumptions
Rejection monitoring	Reads donor DNA fragments in blood	Bloom 2017: NPV 96% for antibody-mediated rejection at a 1.0% cutoff	Positives are often false alarms
Organ viability	Uses perfusion data to judge marginal organs	Perfusion RCT: 50% lower discard in liver (2018)	Trial tested the hardware, not a prediction model
Equity auditing	Measures who benefits under a rule	KDPI race analysis (2023): high-KDPI share for Black donors from 31.09% to 17.75%	Needs a human decision to act

Misconceptions about AI and organ matching

"The shortage is only about supply." The shortage is real, since there are more patients than organs, but the kidney data show a second problem. More than a quarter of recovered kidneys were not used in 2023. Better decisions can recover some of them, though not all, since some organs are truly unusable.

"A more accurate model is a fairer model." Accuracy and fairness are different properties. The KDPI analysis shows a variable that barely changed accuracy while moving a large share of one group's organs across a threshold. A model can be accurate for a population and unfair to a subgroup at once.

"AI picks who lives and who dies." Allocation is a ranking under rules that people write. A model can estimate risk and benefit. A committee decides how much weight each deserves. The ethical choice sits in the weights.

"Simulation proves it works." A simulation proves that a rule would work in a world that behaves like the simulator. It is a necessary step and an insufficient one. Prospective pilots are what separate a promising model from an adopted one.

"The immune match is solved." HLA matching is one input among several. Antibody history, prior transplants, and drug levels all affect rejection risk, and the immune system is far from fully predictable (how the immune system works).

What happens before and after the organ moves

A model that improves allocation touches a small part of the transplant journey. Before an offer, a patient must be evaluated, listed, and kept healthy enough to receive an organ, and kidney patients often wait on dialysis for years, a burden shaped by how well the failing organ's job can be replaced (how the kidneys work). The 2023 kidney report found that 11.4% of adult candidates on the deceased-donor list had waited five years or more, which suggests how much of the problem is queue time and not just matching.

After the transplant, the patient takes immunosuppressant drugs for life, balancing rejection against infection and cancer risk. Dosing is a second candidate for personalized prediction, and the 2026 lung transplant review identified personalized immunosuppressant dosing among the more mature AI applications. Transplant volume itself also varies widely between countries, shaped by donation laws, infrastructure, and culture (how countries compare on organ transplants).

FAQ

Q: How does the US organ allocation system currently work? A: The OPTN, long operated by UNOS under federal contract, maintains waiting lists and separate matching rules for each organ. When an organ becomes available, the system ranks compatible candidates using factors such as blood type, immune match, geography, waiting time, and medical urgency. Transplant centers then work down the list until an offer is accepted.

Q: Can AI predict transplant rejection? A: Partly. Blood tests for donor-derived cell-free DNA can flag injury, and in a 2017 study of 102 kidney recipients a 1.0% cutoff had a negative predictive value of 96% for antibody-mediated rejection. Machine learning can combine that signal with other data, but biopsy remains the reference test in many cases.

Q: Can machine learning reduce kidney discard? A: Possibly, but the evidence for AI specifically is thin. Nonuse was 27.9% in 2023, and the highest rates were in older-donor and high-KDPI kidneys. Better benefit-over-alternatives prediction, perfusion hardware, and protocol changes, as seen with hepatitis C donors, are all plausible routes.

Q: Are any AI allocation tools approved for routine clinical use? A: Reviews do not describe a machine learning model that has replaced a national allocation algorithm. They describe a field of retrospective studies and simulations, and the 2026 liver review argues that AI should support rather than replace clinicians. Tools for monitoring and viability are closer to practice than allocation itself.

Q: Could an algorithm make allocation less fair? A: Yes, if it learns from biased records or optimizes one goal and ignores others. The 2023 KDPI analysis showed how a single race variable shifted the share of Black donors' kidneys above the high-KDPI threshold from 31.09% to 17.75% when removed. Fairness audits and human oversight are the proposed safeguards.

Q: Is the problem a shortage of organs or a shortage of decisions? A: Both. Demand exceeds supply, which no algorithm can change. But in 2023 more than a quarter of recovered kidneys went unused, so some of the gap is about acceptance, logistics, and risk assessment, which are decision problems.

The takeaway

The most useful thing AI has done for transplantation so far is make trade-offs measurable. It can estimate how many deaths a liver rule might prevent, how fast a kidney policy can be tested, and which groups gain or lose when a variable is removed. It cannot decide how much urgency should count against benefit, or whether a woman on a waiting list should receive a different weight than the score gives her. The 27.9% nonuse figure, the 31.09% to 17.75% shift in the KDPI analysis, and the 50% lower discard rate with warm perfusion all point to the same conclusion: the gains are in better decisions about organs that already exist, and those decisions belong to people who can explain them.

---

Sources checked for this update

• OPTN/SRTR 2023 Annual Data Report: Kidney, American Journal of Transplantation, 2025
• Bertsimas et al., Optimized Prediction of Mortality for liver transplant candidates, American Journal of Transplantation, 2019
• Normothermic liver preservation randomized trial, Nature, 2018
• Bloom et al., cell-free DNA and kidney rejection, JASN, 2017

Read next:

• AI Against Superbugs: Machine Learning's Role in Antibiotic Discovery
• AI Meets CRISPR: How Machine Learning Is Supercharging Genetic Engineering
• AlphaFold Explained: How AI Solved Biology's 50-Year Grand Challenge