A Comparison of Risk Adjustment Models Based on Traditional Statistical and Machine Learning Techniques

• Activity Number: 494 - Statistical Methodologies for Identifying, Modeling, and Managing Subpopulations at Risk
• Type: Topic Contributed
• Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Hong Kan* and Hsien-Yen Chang and Hadi Kharrazi
• Companies: and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins Bloomberg School of Public Health
• Keywords: risk adjustment ; machine learning ; predictive models

Abstract

Risk adjustment models are used by public and private health plans and care delivery organizations to adjust for differences in patient characteristics in predicting healthcare resource use and clinical outcomes. This study evaluates the performance of different risk prediction models using patient characteristics in the previous years to predict future total health care costs with administrative claims aimed to guide model selection. The study compares traditional statistical models (eg, ordinary least squares) with machine learning techniques (eg, lasso, elastic net, ridge, adaptive lasso, random forests, and M5). Model performance is evaluated with R squared, mean squared prediction error, and mean absolute error stratified by predicted cost. This retrospective cohort study uses IMS LifeLink database comprised of commercial health plan data in the U.S. Patients with 5 years of continuous enrollment are included in the study. Patient characteristics are extracted using the Johns Hopkins ACG system.