Predicting Healthcare Costs Using Classifiers

Abstract

In the battle to control escalating health care costs, predictive models are increasingly employed to better allocate health care resources and to identify the “best” cases for preventive case management. In this investigation we predicted the top 0.5% most costly cases for year N+1, given a population in year N, with data for the period 1997-2001 taken from the MEDSTAT Marketscan Research Database for a cohort of privately insured individuals diagnosed with diabetes. We considered two performance metrics: i) classification accuracy, i.e. the proportion of correctly classified persons in the top 0.5% and ii) the total number of dollars associated with the predicted top 0.5% of most costly cases.