To combat algorithmic bias in healthcare, including race and ethnicity is critical, a new study says.
Algorithms are used to make healthcare decisions, and can often be more accurate than a clinical assessment, but also carry the risk of bias and discrimination. The study, published in the latest issue of Health Affairs, emphasized that while a common precaution might be removing race and ethnicity data altogether, “this approach is misguided.” In fact, knowledge, not ignorance, of race and ethnicity helps combat bias, researchers said.
While self-reported data is the gold standard for race and ethnicity, estimating (imputing) this data can help identify and address bias. Existing validated methods, like Bayesian Improved Surname Geocoding (BISG), predict race and ethnicity and assess health equity. They were developed by the RAND Corporation, Carnegie Mellon, Kaiser Permanente and the Centers for Medicare and Medicaid Services to mitigate limitations of Medicare data and estimate the probability that each beneficiary would self-identify as a certain racial or ethnic group.
Value-based strategies, while intended to improve care quality, can exacerbate disparities depending on how they are designed. Providers who perform poorly on measures that influence incentive payouts are more likely to care for disadvantaged patients, and may thus receive smaller reimbursements, making it harder to improve care delivery.
Enforcing algorithmic fairness can help ensure payments are not correlated with the racial and ethnic makeup of their patients. This requires estimating those demographics and aggregating them to the provider level.
Though imputation methods are highly accurate, the study noted, “there will always be some uncertainty about the race and ethnicity of a single person.” As a result, if a provider needs to use an algorithm that requires race and ethnicity to inform a clinical decision for one patient, they should seek self-reported data. When imputed data is aggregated, however, uncertainty around one person has “limited impact” on the final equity-adjusted performance incentive that a provider receives.
Ultimately, though there are concerns about algorithm bias in healthcare, that is not unique to “human biases breed health care disparities that are often recapitulated by algorithms,” the study concluded.
While it is difficult to enforce equity in human decisions, methodological approaches like BISG can help enforce equity in algorithms.