Machine learning could help decrease related deaths by 20 percent.
The number of patients who need liver donations far exceeds the number of healthy organs available for transplant.
To address this discrepancy, professors at the Massachusetts Institute of Technology (MIT) have developed a new model for ranking wait-listed candidates: Optimized Prediction of Mortality (OPOM). OPOM has the potential to decrease deaths by 20 percent, according to a new study published by the American Journal of Transplantation.
The current national standard for ranking patients awaiting liver transplants, Model for End-Stage Liver Disease (MELD), has been used since 2002. The candidates receive preference based on their likelihood to survive the next three months. In order to account for the decreased likelihood of survival due to other diseases, MELD grants exception points to certain candidates, such as patients with cancer.
But MIT Sloan School of Management Professor Dimitris Bertsimas, Ph.D., says MELD is systematically biased. The United Network for Organ Sharing (UNOS), he claims, “has created arbitrary points to favor cancer patients, which has overcorrected the problem.”
Because 75 percent of oncology patients are male, MELD scores tend to favor men. In addition, women tend to have higher sodium levels and lower muscle mass, which leads to higher MELD scores. This results in non-equitable distribution.
Using real-world historical data from past patients, OPOM uses optimal classification trees (OCTs), a state-of-the-art machine-learning method, to rank candidates. The order is based on the likelihood that the patient will either die or become unsuitable for liver transplantation in the near future.
Says Bertsimas: “OPOM fairly addresses the imbalance without biasing liver cancer patients.”
To test OPOM’s efficacy, researchers used the most recent Liver Simulation Allocation Model (LSAM), which pulled data from 2007 to 2011. Researchers compared MELD scores with OPOM scores to determine which model was more accurate and efficient.
Results showed that OPOM’s prioritization ranking proved more effective. There was a higher survival rate among all candidates — regardless of where they lived, their demographics or their diagnoses. The new model also distributed livers to a higher number of women.
“Correcting of gender bias and bias for cancer patients are the two most important aspects of OPOM, second only to saving 400 lives a year,” Bertsimas says.
There is no evidence of the large-scale financial benefits of OPOM, but he says that there are significant financial benefits to decreasing the death rate.
Similar state-of-the-art machine-learning models can be adapted to improve the mortality rate for other organ transplants wait-lists, as well. “We are currently working on kidney patients and down the road heart and lung transplant patients,” Bertsimas adds.
The next step is replacing MELD with OPOM as the national standard for ranking liver transplant candidates. The MIT team will present its work to overseers this spring.
In addition to being transparent and accessible for doctors, Bertsimas says that OPOM “is accurate, it is data-driven, and it saves 400 lives a year. I believe it is ready for prime time.”
Get the best insights in healthcare analytics directly to your inbox.