For three years, the Organ Procurement and Transplantation Network has relied on an AI-assisted scoring model called AlloScore to help determine which patients on the transplant waiting list receive available organs.
The system was marketed as a way to remove human bias from life-and-death decisions. It was supposed to be fairer than the old method, which relied heavily on a physician's subjective assessment of a patient's likelihood of post-transplant survival.
But an analysis of 14,000 transplant decisions made between 2023 and 2025, obtained by Autominous through a Freedom of Information request, reveals a pattern the system's developers at MedTech Dynamics were aware of but did not disclose to the hospitals using their software.
Patients from ZIP codes in the bottom quartile of median household income were scored, on average, 11.3 points lower on the AlloScore survival-probability metric than patients from the top quartile — even when controlling for age, diagnosis, and comorbidities. The scoring gap correlates almost perfectly with insurance type: patients on Medicaid scored lowest.
The reason appears to be a training data problem. AlloScore was trained on post-transplant outcomes data that included readmission rates and medication adherence. Both metrics are strongly correlated with socioeconomic factors — patients who cannot afford follow-up visits or who lack reliable transportation are more likely to be readmitted and less likely to maintain strict medication schedules.
The model learned that poverty predicts worse outcomes. Then it used that learning to deny poorer patients access to organs.
"This is the definition of a feedback loop," said Dr. Amara Osei, a bioethicist at Johns Hopkins who reviewed the data at our request. "The system is not measuring medical suitability. It is measuring privilege and calling it survival probability."
MedTech Dynamics declined to comment on the record. In an email exchange, a spokesperson said the company "continuously reviews its models for fairness and accuracy" and that "all scoring outputs are advisory and final decisions remain with clinical teams."
But interviews with transplant coordinators at six hospitals reveal that AlloScore recommendations are followed in 91% of cases. The "advisory" framing obscures the system's actual influence on who lives and who dies.
The Department of Health and Human Services has opened a preliminary review.
What we know for certain
AlloScore is used by the majority of US transplant centres. FOI data shows a measurable scoring gap correlated with socioeconomic status. MedTech Dynamics was aware of the correlation. HHS has opened a preliminary review.
What we are inferring
The training data methodology made this outcome predictable. The company's decision not to disclose the known bias to hospital clients suggests a prioritisation of commercial relationships over transparency.
What we couldn't verify
Whether any specific patient was denied an organ they would have received under the previous system. MedTech Dynamics declined to provide access to their model weights or full training data methodology.