Potential for Algorithmic Bias in Clinical Decision Instrument Development.

Clinical decision instruments (CDIs) face an equity dilemma. They reduce disparities in patient care through data-driven standardization of best practices. However, this standardization may perpetuate bias and inequality within healthcare systems. We perform a quantitative, systematic review to characterize four potential sources of bias in the development of 690 CDIs. We find evidence for potential algorithmic bias in CDI development through various analyses: self-reported participant demographics are skewed-e.g. 73% of participants are White, 55% are male; investigator teams are geographically skewed-e.g. 52% in North America, 31% in Europe; CDIs use predictor variables that may be prone to bias-e.g. 1.9% (13/690) of CDIs use Race and Ethnicity; outcome definitions may introduce bias-e.g. 26% (177/690) of CDIs involve follow-up, which may skew representation based on socioeconomic status. As CDIs become increasingly prominent in medicine, we recommend that these factors are considered during development and clearly conveyed to clinicians.