<p>Precision recall curve.</p>

Maternal mortality in Tanzania remains a public health crisis, with Hypertensive Disorders of Pregnancy (HDP) causing 34% of direct obstetric deaths. In overburdened government clinics, high patient volumes and limited resources often restrict assessments to single-point blood pressure checks, leading to missed diagnoses. This study investigates the potential of machine learning (ML) to move beyond simple threshold detection toward automated risk stratification, aiming to optimize patient flow and prioritize clinical resources for high-risk individuals. We analyzed 337,027 routine records (2023–2024) from Tanzania’s Unified Community System (UCS). Data from multiple visits were aggregated into 187,438 unique client records. HDP was defined by standard clinical thresholds (BP ≥ 140/90 mmHg). We trained five ML models on a balanced subset and validated the top performer on an independent dataset of over 120,000 records to evaluate its utility as a triage tool. XGBoost was the best performing model, achieving 90.1% accuracy and an AUC of 0.95. The model maintained 100% sensitivity, successfully stratifying 12,603 clients into the high-risk category, including those potentially overlooked by traditional checks. While precision was 14% (representing 6.3 false positives per true case), this high-sensitivity screening approach ensures no at-risk client is missed, allowing providers to focus intensive assessment time where it is most needed. ML-driven risk stratification can transform congested ANC workflows by identifying high-risk clients before they escalate to critical states. By automating the initial triage, health facilities can improve operational efficiency and ensure limited specialist time is dedicated to the most vulnerable patients. We recommend that the Ministry of Health strengthens digital data integration to support the deployment of these stratification tools within routine primary care.