Machine learning for risk stratification of hypertensive disorders of pregnancy:
July 02, 2026
Machine Learning for Risk Stratification of Hypertensive Disorders of Pregnancy: Enhancing Clinical Efficiency in Low-Resource Antenatal Care in Tanzania
Study Overview
This study investigated how machine learning (ML) can improve the early identification of Hypertensive Disorders of Pregnancy (HDP) using routine antenatal care (ANC) data collected through Tanzania's Unified Community System (UCS). HDP contributes to approximately 34% of direct maternal deaths in Tanzania, yet routine ANC in many health facilities relies on single blood pressure measurements that can miss women at high risk.
Researchers analyzed 337,027 routine ANC records collected between 2020 and 2024, representing 187,438 unique pregnant women across 23 regions of Tanzania. Five machine learning algorithms were evaluated, with XGBoost demonstrating the best performance for identifying women requiring additional clinical assessment.
Rather than replacing clinical diagnosis, the model was designed as a clinical triage tool to help healthcare providers prioritize women who need further evaluation in busy, resource-constrained antenatal clinics.
Key Findings
The XGBoost model achieved an overall 90.9% accuracy with an Area Under the Curve (AUC) of 0.95, demonstrating excellent ability to distinguish women at risk of hypertensive disorders.
Most importantly, the model achieved 100% sensitivity, meaning it successfully identified every woman classified as having HDP in the validation dataset. This "safety-first" approach minimizes the risk of missing high-risk pregnancies, although it intentionally generates more false-positive alerts that require follow-up assessment.
Blood pressure measurements (systolic and diastolic) were the strongest predictors of HDP, while additional variables—including body mass index (BMI), proteinuria, blood glucose, temperature, and syphilis status—improved overall risk stratification by capturing broader maternal health profiles.
The study also highlighted important health system challenges. Routine health data contained substantial missing laboratory information, with blood glucose missing in 76% of records and several diagnostic variables poorly documented. Despite these limitations, machine learning remained effective when trained on real-world government health data.
Researchers found that many women attended only one ANC visit, making longitudinal clinical monitoring difficult. The ML model therefore provides an opportunity to improve clinical decision-making even when only limited patient information is available.
Conclusion
This study demonstrates that routine digital health data collected through Tanzania's Unified Community System can successfully support machine learning models for maternal risk stratification in low-resource settings.
Rather than replacing healthcare providers, the model serves as an intelligent screening tool that helps prioritize women requiring additional clinical assessment, enabling more efficient use of limited healthcare resources while reducing the likelihood of missed hypertensive disorders during pregnancy.
Before national implementation, the authors recommend strengthening digital health infrastructure, improving data quality, integrating the model into routine clinical workflows, and conducting phased pilot studies to evaluate operational impact, workload implications, cost-effectiveness, and equity. With appropriate implementation, machine learning has the potential to improve maternal health outcomes by supporting earlier identification and management of high-risk pregnancies across Tanzania.
The paper was published in 2026. It appeared in PLOS Digital Health and examined the application of machine learning for identifying hypertensive disorders of pregnancy using routine antenatal care data collected through Tanzania's Unified Community System.