Predicting quality measure completion among 14 million low-income patients enrolled in medicaid.

Abstract

Low-income populations have disproportionately low completion of recommended healthcare services, from missed vaccinations to cancer screenings. While machine learning models help identify high-risk patients for targeted treatment, they have rarely been evaluated for quality measure gap completion-or among low-income populations underrepresented in typical datasets. Analyzing 14.2 million Medicaid recipients-including those excluded from electronic health records and without prior utilization-we developed models to predict gaps in nine nationally adopted quality measures, including preventive care and chronic disease management. Using clinical data to prioritize outreach, the clinical-only model improved accuracy by 32.5 percentage points (pp) over non-predictive methods such as alphabetical calling or birthday reminders (AUROC: 0.88, F1-score: 0.69). Incorporating social determinants of health data further improved performance by 2.0pp in accuracy (to 84.5%) and increased F1-score by 5.0pp (to 0.74), with no change in AUROC (area under the receiver operating characteristic curve). Compared to the clinical-only model, the SDoH model also reduced pre-existing Black-White disparities in prediction accuracy. Model performance was especially sensitive to SDoH factors like healthcare workforce and facility availability.