An analytics approach to problems in health care

Abstract

Health care expenditures in the United States have been increasing at unsustainable rates for more than thirty years with no signs of abating. Decisions to accept or reject deceased-donor kidneys offered to patients on the kidney transplantation wait-list currently rely on physician experience and intuition. Scoring rules to determine which end-stage liver disease patients are in most dire need of immediate transplantation have been haphazardly designed and reactively modified in an attempt to decrease waitlist mortality and increase fairness for cancer patients. For each of the above problem settings, we propose a framework that takes real-world data as input and draws upon modern data analytics methods ranging from mixed integer linear optimization to predictive machine learning to yield actionable insights that can add a significant edge over current practice. We describe an approach that, given insurance claims data, leads conservatively to a 10% reduction in health care costs in a study involving a large private US employer. Using historical data for patients on the kidney waitlist and organ match runs, we build a model that achieves an out-of-sample AUC of 0.87 when predicting whether or not a patient will receive a kidney of a particular quality within three, six, or twelve months. Given historical data for patients on the liver waitlist, we create a unified model that is capable of averting an additional 25% of adverse events in simulation compared to current practice without disadvantaging cancer patients.