You Liver and You Learn
The Organ Procurement and Transplantation Network (OPTN) is in charge of allocating livers as they arrive to wait-listed candidates. Consensus among doctors is that a fair policy should prioritize patients whose conditions would deteriorate most quickly without a transplant. Current practice prioritizes patients who are most likely to die or become medically unfit for transplant within 3 months. Sam Gilmour, Josh Wilde, and I used machine learning and optimization to explore candidate ranking policies based on projected post-transplant outcomes rather than pre-transplant vitality.
6.867 Machine Learning
A policy based on post-transplant outcomes necessarily accounts for organ features as well as patient features. We formulated a maximum weighted matching model in the lineage of work by Bertsimas et al. The weights were equal to the gain that the transplant system would receive by matching an organ to a particular patient, quantified by organ acceptance likelihood and projected post-transplant survival. We further constrained the model by including demographic fairness considerations. We constructed a Lagrangean relaxation of the fairness constraints to obtain fairness-adjusted weights in a classical matching problem.
We required a method for which system benefits could be estimated for newly observed patient-organ pairs. We obtained these estimated benefits as the solutions to a standard OLS regression problem, in which patient and organ features predicted fairness-adjusted weights to the above matching problem.
We iterated with a dynamic policy that changes according to the patients on the waitlist. We hypothesized that the fairness-adjusted weights depend on the characteristics of the current patient waitlist and the organs available for transplant. Our strategy was to predict this dependence by regressing over batch-level organ and patient features to predict one fairness-adjusted weight in the static priority map. Collectively, the learned parameters formed a dynamic policy map generator that yielded fairness-adjusted weights for a given batch of patients and organs.
We built a model that simulates liver allocation and survival outcomes to benchmark the performance of our policies against existing ranking policies. Simulations showed that our policies resulted in approximately 1% more transplants and consequently 1% fewer pre-transplant deaths, with minimal effect on post-transplant morbidity and with continued fair allocation for targeted demographics. Furthermore, our dynamic policy performed at par with the static policy, indicating that its additional computational overhead may be unnecessary.