An individual-level socioeconomic measure for assessing algorithmic bias in health care settings: A case for HOUSES index

Abstract

AbstractWhile artificial intelligence (AI) algorithms hold great potential for improving health and reducing health disparities, biased AI algorithms have a potential to negatively impact the health of under-resourced communities or racial/ethnic minority populations. Our study highlights the major role of socioeconomic status (SES) in AI algorithm bias and (in)completeness of electronic health records (EHRs) data, which is commonly used for algorithm development. Understanding the extent to which SES impacts algorithmic bias and its pathways through which SES operates its impact on algorithmic bias such as differential (in)completeness of EHRs will be important for assessing and mitigating algorithmic bias. Despite its importance, the role of SES in the AI fairness science literature is currently under-recognized and under-studied, largely because objective and scalable individual-level SES measures are frequently unavailable in commonly used data sources such as EHRs. We addressed this challenge by applying a validated individual-level socioeconomic measure that we call the HOUSES index. This tool allows AI researchers to assess algorithmic bias due to SES. Although our study used a cohort with a relatively small sample size, these study results highlight a novel conceptual strategy for quantifying AI bias by SES.