Organ Allocation Policy Evaluation

Last Friday, Diwakar Gupta, from the University of Minnesota, came to Dartmouth as part of the Jones Seminars on Science, Technology, and Society series to discuss different methods of data analysis that could lead to efficient organ allocation policies in the United States.

Gupta and his team of analysts focus on data from end-stage organ disease patients for whom organ transplantation is the only treatment available. They found that, under the current system, for every ten transplants, four patients die while waiting to receive an organ. Gupta and his team also discovered that, on average, 21 offers are made for an organ before centers find a recipient willing to take the organ.

To demonstrate the current methodology behind organ allocation, Gupta focused on liver transplant recipients. For these patients, factors ranging from location and age, to model for end-stage liver disease score (MELD), which rates the severity of the disease, factor into the decision. When placing an organ for transplant, authorities first rank by location, with those who are not considered local or regional placed at the end of the list due to complications related to transporting organs. The next factor considered is MELD score. But after these first two criteria, the policy becomes very complicated.

Gupta then discussed the various analytical methods and programs used by his team to evaluate current organ allocation policies. His team focused on machine learning techniques, which identify patterns from data on factors such as age of the donor and recipient, body mass index, and numerous other characteristics.

Gupta favored a specific class of models that used support vector machines. These models take the input data and predict which category, in this case yes for the acceptance of an organ or no for the rejection of an organ, patients will fall into using complex learning algorithms.

According to Gupta, the methods used in support vector machines prevent over-fitting to training data and lead to better separation of organ rejections from organ acceptances. Unfortunately, it can take over a day to run the algorithms used in this analysis, which poses a problem when under strict time constraints.

In addition to evaluating support vector machines, Gupta studied classification and regression trees (CART), which look at all attributes of donors and patients and try to find the attribute with the strongest correlation with a yes decision. He also compared machine learning methods to the naïve method, which equates to coin flipping for a yes or no decision, and found that this method was better at predicting a correct response than all of the machine learning methods he had evaluated.

Gupta and his team are currently compiling their findings and putting them into a paper that they hope will be read by policy makers and analysts. They would like to promote the use of support vector machine models for data analysis and policy formation and, ultimately, save the lives of end-stage organ disease patients.