How analytics and machine learning can aid transplant decisions
Imagine this scenario: A patient named John has waited 5.5 years for a much-needed kidney transplant. One day, he learns that a deceased donor kidney is available and that he is the 153rd patient to whom this kidney was offered.
Clearly, this is not a “high-quality” organ if it was declined by 152 patients. But John has been waiting a long time for a new kidney. Should John accept or decline the kidney? And can analytics and machine learning help make that decision easier?
Currently, that decision is usually made by John’s doctor based on a variety of factors like John’s current overall health status on dialysis and a gut instinct about whether (and when) John will get a better offer for a healthier kidney. If John is young and relatively healthy, the risk of prematurely accepting a lower-quality kidney is future organ failure and more surgeries. If John’s health status is critical and he rejects the kidney, he could be underestimating how long it will take until a higher-quality organ is available. The decision could be a matter of life or death.
John’s dilemma isn’t unique in the world of kidney transplantation, where current demand outpaces supply. Since 2002, the number of candidates on the waitlist has nearly doubled, from just over 50,000 to more than 96,000 in 2013. During the same time, live donation rates have decreased. Complicating this problem of supply and demand is an unacceptably high deceased donor organ discard rate – up to 50% in some instances.
The desire to provide patients with the highest quality organ has the potential to become the doctors’ Achilles heel. In striving to maximize patient outcomes for an individual patient, the discard rate could increase. This crossroads of physicians’ semi-quantitative calculus of a patient’s health factors and “gut instinct” would greatly benefit from a data-driven tool to assist in this complex decision-making process.
With colleagues at MIT Sloan and Massachusetts General Hospital, we developed an analytics tool to help doctors in deceased-kidney acceptance decisions. The model aims to calculate the probability of a patient being offered a deceased-donor kidney of a certain quality level within a specific time frame (three, six, or 12 months), given their individual characteristics. Using machine learning, it looks at 10 years of data and millions of prior decisions to estimate a patient’s waiting time in the context of a current active organ offer until the time to the next offer for a higher quality kidney.
As for accuracy, we tested the model against real outcomes in different states. For example, in California, the actual probability of getting a high-quality kidney in the next six months for a patient with John’s same health factors was 1.3 to 1.7 percent. The model predicted 2 percent. In Maryland, the actual probability was 8 to 16 percent, and our model predicted 10.4 percent. In New York, the actual probability was 3 to 10 percent, and the model predicted 5.4 percent. Overall, the estimation of accuracy (AUC) was 87 percent, which illustrates that the model produces credible predictions.
If widely accepted, this tool could assist practitioners in their organ acceptance decisions, and serve as an educational tool for candidates awaiting transplantation. By defining the future organ offer landscape in a patient-specific format, we hope to not only provide doctors with the ability to achieve expedited, evidence-based decision-making, but also to provide an interactive educational tool for transplant candidates to further their understanding of an additional aspect of the risk/benefit ratio associated with offer acceptance – specifically the factor of additional waiting time.
In other words, this model could have a significant impact by helping to make the entire donation system more efficient. It could produce better matching of organs overall and contribute to fewer discarded kidneys.
The model could also assist other types of transplants like liver transplants. It could be useful in any situation where physicians and patients must make decisions without knowing what the future holds – and have to balance a current offer with potential future offers.
While we are still finalizing our study, we hope to make the tool available for hospitals to use in the near future. Developing an app-form of the tool for doctors and patients is also in our plans.
We are hopeful that when patients like John get an organ offer, they and their doctors will have access to a data-based tool to help make the best decisions possible not only for themselves, but for the entire transplantation network.