Maintaining continuous Medicaid eligibility is essential for ensuring access to healthcare services, continuity of care, and effective administration of public health programs. Despite ongoing policy efforts to improve retention, short-term eligibility loss remains frequent and often occurs with little advance notice. This study presents an interpretable machine learning framework for predicting next-week Medicaid dropout using longitudinal weekly eligibility records. Leveraging 52 weeks of binary eligibility indicators, the problem is formulated as a one-week-ahead prediction task, and transparent temporal features are constructed to summarize recent participation intensity, continuity, and instability. These longitudinal signals are integrated with demographic characteristics, including age group, sex, primary language category, and county of residence. Multiple ensemble learning models are assessed under extreme class imbalance using both discrimination and outreach-oriented ranking metrics, including ROC-AUC, PR-AUC, recall@10%, and lift@10%. The results indicate that incorporating demographic context substantially enhances predictive performance. The strongest model, CatBoost with demographic features, achieves a recall@10% of 0.65 and a lift of 6.5×, supporting effective prioritization for targeted outreach. Overall, the findings highlight the benefit of combining eligibility behavior over time with demographic context to enable proactive Medicaid retention strategies.