There is a challenge to donation after circulatory death, however: time.

While the donor is dying, the blood supply to organs throughout the body can vary and, in some cases, stop altogether, leading to liver damage. The liver contains a network of pipes called ducts that squeeze out bile, a fluid that helps us digest food, to the gallbladder and intestines. A long time between the cessation of life support and the donor’s time of death is associated with malfunctioning ducts and serious complications for transplant recipients. If the donor’s time of death happens more than 30 minutes after blood flow starts to decrease to the body’s organs, the liver might not be useful for transplantation.

About half of the possible donors die within the first 30 minutes after life support is removed. When death occurs later, between 30 and 60 minutes after life support ends, surgeons use their judgment to determine which donors are the best candidates by considering the donor’s vital signs, blood work, and neurological information such as the pupil and gag reflex. Still, about half of the transplantations need to be canceled because death occurred too late. Knowing where to allocate resources, such as normothermic machine perfusion devices, can save money and streamline the workload of transplant health care workers, Sasaki explained.

Competing machine-learning algorithms

To predict the time of death, the model uses an array of clinical information from the donor including gender, age, body mass index, blood pressure, heart rate, respiratory rate, urine output, blood work test results and cardiovascular health history. The model also considers the ventilator settings, which indicate how much help someone needs to breathe, in addition to neurological assessments of how conscious the patient, as well as pupil, corneal, cough, gag and motor reflexes.

The research team pitted numerous machine-learning algorithms against each other to find the one that best predicted the time of death using the same information available to surgeons. The winning algorithm was more accurate than surgeons and other available computerized tools at predicting whether the donor’s time of death would happen within the acceptable time frame for a successful transplant. The model was trained and validated on more than 2,000 real-world cases from six U.S. transplant centers.

The model accurately predicts the donor’s time of death 75% of the time, outperforming both existing tools and the average judgment of surgeons, who accurately predicted the time of death 65% of the time. It also makes accurate predictions for cases with information missing from the medical record.

The research team designed the model to be customizable so it can handle different surgeon preferences and hospital procedures. For example, the model can be set to calculate the time of death from when life support is removed or from when agonal breathing, a gasping breathing pattern that happens as a body is dying, begins. The researchers have also developed a natural language interface, similar to ChatGPT, that pulls information from the donor medical record into the model.

Minimizing missed opportunities

Sometimes death unexpectedly occurs within the time frame when organs are suitable for transplantation, but because preparations must be underway before the donor dies, these cases do not result in a transplant. The rate of these missed opportunities was similar for the model and surgeon judgment: Both were just over 15%.

Because artificial intelligence is rapidly advancing, the researchers expect that the model’s accuracy in predicting time of death will improve and that it will catch more missed opportunities.

“We are now working on decreasing the missed opportunity rate because it is in the patients’ best interest that those who need transplants receive them,” Sasaki said. “We continue to refine the model by having competitions among available machine learning algorithms, and we recently found an algorithm that achieves the same accuracy in predicting the time of death but with a missed opportunity rate of about 10%.”

The research team is also working on variations of the model for use in heart and lung transplants.

Researchers from the International University of Health and Welfare, Duke University School of Medicine, Cleveland Clinic, University of Rochester Medical Center, University of Florida College of Medicine, Virginia Commonwealth University Health, Columbia University Irving Medical Center, and Transmedics, Inc. contributed to this study.