However, donation after cardiovascular death presents a time challenge.
While the donor is dying, the blood supply to organs throughout the body changes and sometimes stops completely, which can lead to liver damage. The liver contains a network of pipes called ducts that express bile, a fluid that helps digest food, to the gallbladder and intestines. The long time between withdrawal of life support and death of the donor results in vascular malfunction and severe complications in the transplant recipient. If blood flow to the organ begins to decrease and the donor dies for more than 30 minutes, the liver may not be useful for transplantation.
About half of potential donors die within the first 30 minutes after life support is removed. Then, if death occurs 30 to 60 minutes after life support is removed, surgeons consider the donor’s vital signs, blood tests, and neurological information such as pupillary and gag reflexes to decide for themselves which donor is the best candidate. Still, about half of transplants have to be canceled because death occurs too late. Knowing where to allocate resources such as normothermic mechanical perfusion equipment can save costs and streamline the workload of transplant providers, Sasaki explained.
Competing machine learning algorithms
To predict time of death, the model uses a range of clinical information from the donor, including gender, age, BMI, blood pressure, heart rate, breathing rate, urine output, blood test results, and cardiovascular health history. The model takes into account a patient’s state of consciousness, neurological assessments such as pupils, corneas, cough, gagging, and motor reflexes, as well as ventilator settings that indicate how much assistance is needed to breathe.
The research team pitted a number of machine learning algorithms against each other to find the one that best predicted time of death using the same information available to surgeons. The winning algorithm was more accurate than surgeons and other available computerized tools in predicting whether a donor’s time of death would occur within an acceptable time frame for transplant success. The model was trained and validated on more than 2,000 real-world cases from six transplant centers in the United States.
The model accurately predicted the donor’s time of death 75% of the time, outperforming both existing tools and the surgeon’s average judgment, which correctly predicted the donor’s time of death 65% of the time. It also accurately predicts cases where information is missing from the medical record.
The research team designed a customizable model to accommodate different surgeon preferences and hospital procedures. For example, you can set the model to calculate the time of death from when life support is removed or from the beginning of agony breathing, the gasping breathing pattern that occurs when a person is dying. The researchers also developed a natural language interface similar to ChatGPT that feeds information from the donor’s medical records into the model.
Minimize opportunity loss
Sometimes an unexpected death occurs when an organ is suitable for transplantation, but this does not result in a transplant because preparations must be made before the donor dies. The rates of these missed opportunities were similar for model and surgeon judgment, both just over 15%.
As artificial intelligence advances rapidly, researchers hope that models that predict time of death will become more accurate and capture more missed opportunities.
“We are now working to reduce the lost opportunity rate because it is in the best interest of the patient that those who need a transplant receive a transplant,” Sasaki said. “We continue to improve our model by competing among available machine learning algorithms, and recently discovered an algorithm that achieves the same accuracy in predicting time of death, but with an opportunity loss 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 Human Services, Duke University School of Medicine, Cleveland Clinic, University of Rochester Medical Center, University of Florida College of Medicine, Virginia Commonwealth University School of Health, Columbia University Irving Medical Center, and Transmedics, Inc. contributed to this study.
