ML Model Achieves 100% Sensitivity for Pregnancy Hypertension Screening in Tanzania

In Tanzania, where hypertensive disorders of pregnancy account for roughly one-third of all maternal deaths, clinics face an impossible arithmetic: too many patients, too few resources, and blood pressure checks that happen once, if at all. A research team led by investigators at Prime Health Initiative Tanzania set out to ask whether a machine learning model could do better—not by replacing clinical judgment, but by catching cases that a single snapshot of numbers might miss.

The researchers worked with 337,027 antenatal care visits recorded between 2020 and 2024 in Tanzania's Unified Community System, a digital platform for health data collection across 23 regions. When they collapsed these visits into individual patient records, they had 187,438 women to work with. The data told a story of scarcity: the median woman had attended just one clinic visit. Only 6.3 percent had reached four visits. Almost none—fewer than one in a thousand—had achieved the WHO-recommended eight visits. This sparse reality shaped everything that followed. The researchers could not build a model that relied on tracking changes over time. They had to work with what most clinics actually had: a single encounter, often incomplete, often recorded hastily.

They trained five different machine learning algorithms on a balanced subset of their data, then tested the best performer—an XGBoost model—on an independent validation set of over 120,000 records from April and May 2024. The model achieved 90.1 percent accuracy and an area under the curve of 0.95, a measure of its ability to discriminate between high-risk and low-risk cases. But the headline number that mattered most was this: the model identified every single high-risk patient. One hundred percent sensitivity. In a clinic where missing a case of hypertensive disorder could mean a woman's death, that zero-miss rate was the whole point.

The cost of that perfection was precision. The model flagged 12,603 women as high-risk in the validation set. Of those, 1,725—about 14 percent—actually met the clinical threshold for hypertensive disorder (blood pressure of 140/90 mmHg or higher). The other 10,878 were false positives. This meant roughly 6.3 false alarms for every true case. In a busy clinic, that translates to dozens of extra workups per clinician per day. Yet when the researchers looked closely at these false positives, they found something interesting: these women had mean blood pressures of 136.5/84.2 mmHg, compared to 114.4/69.5 mmHg in women who tested negative. They were not random errors. They were women whose blood pressure was elevated but not yet at the diagnostic threshold—a subclinical risk group that a single-point threshold check would have missed entirely.

The model's predictions rested on seven clinical variables: systolic and diastolic blood pressure (which dominated the decision-making), body mass index, blood glucose, proteinuria, temperature, and syphilis status. Blood pressure alone could have done most of the work, but the researchers retained the other variables to capture a fuller biological picture and to ensure the tool would function across the diverse presentations and resource constraints of different clinics. The data itself was messier than the clean datasets used in most published machine learning studies. Nearly three-quarters of the original records were excluded because they lacked complete information—a reflection of real health systems where diagnostic tests are sometimes unavailable or skipped in resource-constrained facilities. The researchers chose not to artificially fill in missing values, because doing so would have created a false sense of precision and masked the uncertainty that actually exists in routine clinic data.

The authors acknowledge the central tension in their work: because the outcome (hypertensive disorder) was defined by the same blood pressure thresholds that served as the model's primary predictors, the model is not truly predicting future disease. It is, instead, demonstrating an enhanced ability to detect consistent patterns of blood pressure elevation in noisy, real-world data—to screen in cases that a human clinician performing a single-visit check might overlook. This distinction matters for how the tool would actually be used. It is not a crystal ball. It is a triage mechanism, a way to say: this woman warrants a closer look, a recheck, a conversation with a specialist. The researchers recommend that implementation include tiered risk outputs (low, moderate, high), same-day nurse rechecks, and low-threshold confirmatory protocols rather than automatic referrals. They also call for continuous local recalibration as more outcome data accumulate, and for prospective pilots to measure whether the gain in sensitivity—catching every case—justifies the operational burden of managing false positives in the real world. Before any nationwide rollout, they argue, Tanzania's Ministry of Health should invest in stronger digital data integration and test the model's impact on actual clinical workflows, resource use, and equity across different facility types.