Development and evaluation of neighborhood social risk indices for surgery using outcome-specific machine-learning models.

Abstract

Social risk factors, which influence health outcomes, are used in measures applied to federal payment adjustment models and in health disparities research for risk adjusting outcomes. However, existing composite measures of social risk have notable limitations in predicting specific health outcomes. Using data from the American College of Surgeons National Surgical Quality Improvement Program hospitals, this study constructed machine learning-based social risk indices, comprised of the underlying US Census data components used in the Area Deprivation Index, a tool developed to measure neighborhood-level disadvantage. Social risk components were tuned to surgical outcomes to assess how social risk influences specific outcomes. In this study of 3,206,836 patients from 688 US hospitals, models using machine learning-based indices outperformed models using the Area Deprivation Index. Machine learning-based models had greater predictive power for all 14 outcome models than models using the Area Deprivation Index directly (median 8.15-fold increase in standardized parameter estimates). Mean calculated contributions of each underlying Area Deprivation Index component used in the machine learning-based indices varied across the 14 outcomes, providing insight into social risk factors that contributed the most to specific postoperative outcomes. The machine learning-based indices developed in this study provide social risk assessments with higher informational value and may function as a more reliable measure in health equity risk adjustments than existing tools currently used for health care payment adjustments.