AI in organ transplantation is so far a research field, not a routine tool. A 2025 systematic review of 20 studies on machine learning for liver transplant allocation found that most predicted survival after transplant, that few urgency models clearly beat the standard MELD score, and that the work was retrospective. A deep learning model for reading donor kidney biopsies outperformed on-call pathologists on one measure in a single 83-kidney study, but that study did not measure real discard decisions.
Key takeaways
- Canada had 4,044 people waiting for a solid organ at the end of 2024, and 70.1 percent of them were waiting for a kidney.
- The 2025 liver review included 20 papers, eight of them from one Spanish team, and none built a model of transplant survival benefit.
- A donor kidney biopsy model reached a correlation of 0.916 with pathologist annotations, against 0.884 for on-call pathologists, in one 83-kidney dataset.
- The reported 37 percent drop in erroneous kidney discards was a modeled estimate, not an observed result.
- No source reviewed shows AI deciding who receives an organ; allocation rules stay with national organ allocation systems and transplant teams.
The shortage that makes prediction attractive
The Canadian Institute for Health Information (CIHI) reports that 3,203 organ transplants were performed in Canada in 2024, 6.3 percent fewer than the 3,417 in 2023 and the first decline in five years. On December 31, 2024, 4,044 Canadians were waiting for a solid organ transplant, and 691 patients died or were withdrawn from the list during the year. There were 895 deceased donors, down 5.9 percent, and 591 living donors. The figures come from CIHI's organ donation and transplantation release of December 18, 2025.
With supply that tight, any method that predicts who is most likely to die while waiting, or which donor organs are safe to use, draws research interest. Prediction is the main thing machine learning does in this field.
Liver allocation: what 20 studies found
Lisiane Pruinelli and colleagues searched the literature from 2002 to June 2023 and reviewed 20 papers on machine learning in liver transplant allocation, published in BMC Medical Informatics and Decision Making in February 2025. The open-access review notes that urgency, utility, and benefit guide allocation, yet allocation itself follows an urgency model. Most included studies predicted survival after transplant. Only two targeted urgency, meaning waitlist death or dropout, and few such models clearly outperformed MELD. No study produced a transplant survival benefit model, and eight of the 20 came from a single Spanish group.
The authors list further limits: reliance on registry data that lack detail over time, little external validation, inconsistent comparison with a baseline model, poor reporting of missing data, and no study that addressed algorithmic bias or fairness. They describe the included work as retrospective and name prospective validation as a future need.
Donor kidney biopsies: a pathology example
Before a deceased-donor kidney is accepted, a pathologist may examine a frozen biopsy section to estimate the percentage of glomeruli, the kidney's filtering units, that are scarred. Marsh and colleagues built a deep learning model to measure this on whole-slide images and reported it in JAMA Network Open in 2021. They used 98 frozen and 51 permanent sections from 83 deceased-donor kidneys, annotated by three board-certified pathologists. The analysis used a 20 percent scarring cutoff as the discard threshold.
The model's correlation with the annotations was 0.916 for single slides, against 0.884 for on-call pathologists, and 0.933 when several levels from one kidney were pooled. A pathologist who reviewed and corrected the model's output did best, with a correlation of 0.958 in a 25-case workflow. The study estimated that pooling four levels could lower erroneous discards from 31 to 13 in a simulation of 1,000 kidneys. That is a projection from prediction intervals. Real discard decisions were not measured, the data came from one center, and the authors call for testing on slides from other laboratories and scanners. Several authors disclosed royalty or financial interests in the licensed technology.
How strong is each piece of evidence?
| Application | What exists | What is missing |
|---|---|---|
| Liver allocation models | 20 retrospective studies, mostly on survival after transplant | Benefit models, outside validation, fairness testing, prospective studies |
| Donor kidney biopsy reading | One 83-kidney model study with a pathologist comparison | Other laboratories and scanners, measured effect on discards |
The table covers only the two bodies of work checked for this article. Other organs and other uses, such as heart or lung matching, were not verified and are left out.
Fairness, governance, and who decides
Allocation is a rule-making problem as much as a prediction problem. A model trained on past registry data learns from past decisions and past gaps in those records, which is why the absence of any fairness analysis in the liver review matters. The World Health Organization's 2021 guidance, Ethics and governance of artificial intelligence for health, published on June 28, 2021, sets out six consensus principles for AI in health. In practice, listing, organ acceptance, and treatment remain decisions for transplant teams working under national allocation policy.
This article is general information about research and statistics. It is not medical advice and does not tell anyone how to approach donation, listing, or care. Questions about a particular case belong with a transplant program or another qualified clinician.
FAQ
Does AI decide who gets a donated organ?
Nothing in the sources reviewed says so. The liver review describes current allocation as an urgency-based system, with MELD 3.0 prioritizing urgency and not accounting for survival after transplant. The machine learning studies it covers are research models.
Could AI reduce the number of kidneys that are discarded?
It might, but the evidence is modeled rather than observed. The 2021 biopsy study projected fewer erroneous discards in a simulation and said clinical impact would need prospective evaluation.
Why do so many studies come from the same groups?
Transplant data are limited and often held in national registries or single hospitals. In the liver review, eight of 20 papers came from one Spanish team, so a few datasets shape much of the literature.
Where do the Canadian transplant numbers come from?
They come from CIHI, which reports annual organ donation and transplantation statistics. The figures here are for 2024, released in December 2025, and a newer release may change them.
Related reading
- How AI is tested in medical diagnosis
- Computer vision and its applications
- Deep learning applications explained
- A review of recent advances in AI
More in AI in health.
Last verified: October 2026
Sources
- Organ donation and transplantation data and reporting, Canadian Institute for Health Information, December 18, 2025
- Pruinelli L et al., Transforming liver transplant allocation with artificial intelligence and machine learning: a systematic review, BMC Medical Informatics and Decision Making, 2025
- Marsh JN et al., Development and validation of a deep learning model to quantify glomerulosclerosis in kidney biopsy specimens, JAMA Network Open, 2021
- Ethics and governance of artificial intelligence for health, World Health Organization, June 28, 2021