Socioeconomic bias in applying artificial intelligence models to health care

Socioeconomic status (SES) is a key dimension along which artificial intelligence (AI) models can be “biased”, or more specifically, along which AI models can exhibit disparate performance across demographic subgroups. However, measuring SES in ways usable for healthcare research is a challenge. We present the HOUSES index, an extensively validated way of measuring SES for healthcare applications, and show its use in detecting and measuring performance disparities in AI models. Going beyond measurement and theorizing about causal mechanisms, we take an understanding of AI as a form of causality-agnostic statistical modeling that automatically finds optimal correlations. With this, we present a hypothesis and supporting evidence that a lack of healthcare access among those of lower SES propagates through to these patients having lower-quality healthcare data, which leaves less of a valid signal for AI models to pick up, ultimately resulting in lower performance for those of lower SES.