Decomposition of Clinical Disparities with Machine Learning

Differences in average rates of access to quality care, mortality from specific diseases, and surgery for conditions such as emergency cardiac care point to racial disparities in healthcare. The optimal approach to alleviate a given disparity depends on whether the main driver is differential health risk or differential treatment within the healthcare system. In this paper, I propose an extension of the Oaxaca-Blinder decomposition framework that capitalizes on advances in clinical data and machine learning prediction to quantify the portions of a given disparity due to differential clinical health and differential healthcare treatment. The proposed method applied to the surgery decision for heart attacks using electronic medical records data from a major academic hospital system in Indiana suggests a smaller potential healthcare treatment disparity compared to the conclusion from the standard approach. The method reveals that 1/3 of the cardiac surgery rate difference can be explained by differences in clinical health variables between Black and non-Black patients pointing towards the existence of worse relative social health risks for patients clinically recorded as Black. Differential health risks for the socially constructed concept of race indicates the need for society-wide solutions to address differences in risk factors such as healthcare access and socioeconomic status. However, a substantial cardiac surgery disparity, constituting 2/3 of the rate difference, remains even after machine learning-based clinical health adjustment pointing towards the need for solutions that target differential clinical treatment.