Practical considerations for social determinant-based disease prediction in the All of Us research program

Abstract

Growing recognition that social determinants of health (SDoH) strongly influence health outcomes has expanded their inclusion in biomedical research, underscoring the need to evaluate how best to incorporate these data into disease prediction models. The All of Us (AoU) Research Program is a large, diverse biomedical research dataset that includes participants from across the United States and links electronic health records (EHRs) with extensive survey data covering a wide range of health, lifestyle, and social factors. We assessed selection bias in the SDoH surveys by comparing demographic characteristics across cohorts with varying EHR and survey completion requirements. We additionally used a series of logistic regression models to evaluate the predictive utility of SDoH for nine chronic conditions, compared these results to models using only socioeconomic status (SES), self-reported race and ethnicity, or additional area-level SDoH factors, and discussed the associated trade-offs. Here we show that requiring sufficient individual-level SDoH survey data results in significant selection bias and sample reduction in AoU. We also show that SES alone captures a substantial proportion of the predictive signal from individual-level SDoH data while preserving sample size and mitigating selection bias. Moreover, SES measures provide greater predictive utility than self-reported race and ethnicity, without excluding underrepresented groups. We find disease-specific patterns of association with SDoH and that area-level SDoH metrics contribute to disease prediction independently of individual-level measures. Altogether, we emphasize key analytical considerations and disease-specific trade-offs for the integration of SDoH data into disease prediction models in AoU and similar cohorts.