MS Speaks

SUMMARY AND COMMENT | GENERAL MEDICINE

RACIAL BIAS IN ALGORITHMS USED TO IDENTIFY PATIENTS WHO NEED SPECIAL ATTENTION


Anthony L. Komaroff, MD--reviewing Obermeyer Z et al. Science 2019 Oct 25

One large health system's algorithm was less likely to recommend interventions for black patients at any level of illness severity.

Many U.S. health systems use proprietary algorithms that analyze electronic medical record data to identify patients who are particularly at risk for adverse health outcomes in hopes of intervening to prevent those outcomes. To examine potential racial disparities in algorithms' recommendations, investigators obtained a medical record dataset from a large health system that covered more than 100,000 patient-years. The dataset included elements used to predict outcomes and the outcomes themselves. Investigators also were given the algorithms' inputs (e.g., demographics not including race, insurance type, diagnosis and procedure codes, medications, and detailed costs) and outputs.



The algorithm assumed that patients who were projected to have the greatest future cost had the greatest need for intervention. However, when self-reported race was taken into account, black patients were sicker than white patients by measures other than cost (e.g., greater number of chronic illnesses) and generated fewer costs at any level of illness severity — because they received less care. Thus, at any level of illness severity, the algorithm was more likely to recommend preventive interventions for white patients than for black patients. The algorithm-maker confirmed these findings on a much larger dataset.

COMMENT

Electronic record databases could be powerful tools to identify patients who need preventive interventions. But these findings argue that, if analytic techniques are suboptimal, programs designed to protect patients might, instead, be harmful to some.


From NEJM Journal Watch (December 10, 2019).