Machine Learning Approaches to Racial/Ethnic Differences in Social Determinants of Mild Cognitive Impairment and Its Progression to Dementia in the All of Us Research Program

Abstract Objective This study examines how social determinants of health (SDOH) influence mild cognitive impairment (MCI) and its progression to dementia across racial/ethnic groups, identifying disparities and key predictors using machine learning approaches. Materials and Methods We analyzed data from 83,180 participants aged 50+ in the All of Us Research Program (65,582 White, 6,207 Black, 4,170 Hispanic, 7,221 Other). The sample had mean ages ranging from 62.4 (Hispanic) to 68.1 (White) years, with significant gender disparities (70.9% Black females vs 46.0% Other females). We developed machine learning classification models to predict MCI and its progression to dementia across the four racial/ethnic groups using 18 SDOH, along with key sociodemographic variables. We then applied SHapley Additive exPlanations (SHAP) to quantify each factor’s contribution and interpret its risk and protective effects on individual predictions. Results MCI prevalence was comparable across groups (7.5-8.0%), but progression to dementia varied (9.4% Black vs 11.4% Other). Perceived stress was the strongest predictor of MCI across all groups, with SHAP values of 15.1% (White), 13.5% (Black), 17.4% (Other), and 19.3% (Hispanic). Predictors of progression to dementia varied by groups: perceived stress (7.0%) for Whites, instrumental social support (14.2%) for Hispanics, daily spiritual experience (34.0%) for Blacks, and everyday discrimination (11.2%) for other groups. Discussion The findings underscore the need for group-specific interventions addressing stress mitigation for MCI prevention and culturally-tailored support systems to delay dementia progression. This machine learning approach reveals complex SDOH interactions that traditional methods might overlook, particularly for racial/ethnic underrepresented populations.