East Coast Ballroom
Can Machine Learning Reduce the Burden of Health Care Quality Measurement? (306678)
Michael E Chernew, Harvard Medical SchoolBruce E Landon, Harvard Medical School
Mary Beth Landrum, Harvard Medical School
J. Michael McWilliams, Harvard Medical School
*Christina A Nguyen, Massachusetts Institute of Technology
Keywords: quality measurement, quality performance, machine learning, accountable care organizations
Quality measurement programs are important for ensuring that quality of care is not suffering, but require substantial investments of time and resources by providers. We examine whether machine learning (ML) could identify low-quality accountable care organizations (ACOs) as defined across the full measurement set using an abridged (low provider burden) set of measures. We used 2014-2017 CMS performance data on 1,284 Medicare Pioneer and Shared Savings Program ACOs. We classified ACOs as low-quality (bottom quartile based on CMS methodology and full set of measures) using an ensemble learning algorithm with bagging, least absolute shrinkage and selection operator estimates, and class weighting to avoid overfitting. ML using the full set correctly classified 78.0% of ACOs (sensitivity=83.6%, specificity=61.5%). Although ML with the abridged set increased classification accuracy only slightly (73.1% using CMS method versus 75.1% using ML), it substantially increased sensitivity (from 46.9% to 78.2%) and decreased specificity (from 81.9% to 65.8%). ML using an abridged set can optimize identification of low-quality providers, and resources could target improving their quality of care.