<p>Maternal and infant mortality remain disproportionately high in the United States compared to other high-income nations, particularly among Medicaid recipients, who account for nearly half of all U.S. births. While the association between social determinants of health (SDoH) and adverse maternal outcomes is well-established, implementing this knowledge in real-time risk stratification has remained challenging. We developed and validated machine learning models that integrate healthcare and SDoH data from 190,698 Medicaid-enrolled women across 26 states and Washington, DC, to support earlier, actionable risk identification in clinical practice. A model using only demographic and clinical data achieved 86.3% accuracy, 93.1% AUC, and 71.3% sensitivity in predicting adverse pregnancy outcomes. Incorporating SDoH—particularly healthcare access variables such as provider availability, distance to care, and infrastructure—improved sensitivity to 81.3% (a 10.0 percentage point gain) while maintaining high specificity (94.3%) and eliminating algorithmic sensitivity disparities between Black and White patients. The model identified risk a median of 55 days before traditional clinical indicators emerged, providing a substantial window for proactive, community-based intervention. Simulation of targeted SDoH improvements, particularly in maternal healthcare workforce availability and infrastructure, predicted a 31.8% reduction in adverse pregnancy outcomes, with the greatest absolute benefit for Black women. These findings suggest that systematically integrating clinical and social data has the potential to identify high-risk pregnancies months before complications emerge, which could enable earlier intervention within existing Medicaid care management frameworks and help address persistent racial inequities in maternal health outcomes.</p>