Predicting Patient Costs

• Activity Number: 554
• Type: Topic Contributed
• Date/Time: Wednesday, August 3, 2016 : 10:30 AM to 12:20 PM
• Sponsor: Scientific and Public Affairs Advisory Committee
• Author(s): Grace Shrader* and Jonathan Berthet and Katherine Tong and David O. Meltzer
• Companies: University of Wisconsin - Madison and The University of Chicago and The University of Chicago and The University of Chicago
• Keywords: Electronic Health Records ; Machine Learning ; Predictive Analytics

Abstract

A more accurate patient cost prediction model would facilitate the development of a prospective payment system in healthcare that incentivizes long-term efficiency and effectiveness of care. Patient cost prediction models currently used by healthcare payers are based in linear regression on only a small number of patient demographic and diagnosis variables. This work introduces a methodology for building patient cost prediction models that (1) extract predictive information from extensive electronic health records, (2) identify medical variables prognostic of a patient's future costs, and (3) improve upon the predictive accuracy of linear regression models through the use of shrinkage methods and tree-based machine learning techniques. Cost prediction models were trained, tested, and evaluated for accuracy on a set of patient data from over 30,000 University of Chicago Hospital patients.