Early Detection of Hypertensive Disorders of Pregnancy in Sub-Saharan African Women: A Multi-Model Machine Learning Approach

Hypertensive disorders of pregnancy (HDP) account for approximately 16% of maternal deaths in Sub-Saharan Africa (SSA), yet decision-support tools calibrated for resource-limited African clinical settings remain scarce. This paper presents MRC-HDP (Maternal Risk Classifier for HDP), a multi-model machine learning framework for early HDP risk stratification in pregnant women. Using the UCI Maternal Health Risk dataset (n = 1,014, three risk classes), we train a Random Forest Classifier with rule-based parametric data augmentation and SMOTE balancing, achieving 83.36% accuracy, Macro F1 of 0.8373, Cohen's Kappa of 0.7487, and MCC of 0.7544 on held-out real data. Five-fold cross-validation yields 78.54% (±2.19%). We additionally present two proof-of-concept temporal models on augmented longitudinal data - an RF Regressor for time-to-event prediction and an LSTM for week-by-week risk trajectory modelling - explicitly framed as architectural demonstrations pending real longitudinal data. A systematic ten-configuration ablation on the 1,014-patient dataset, incorporating the engineered feature Pulse Pressure (PP = SBP − DBP) motivated by its clinical role as an arterial stiffness proxy, demonstrates that Blood Sugar is the dominant predictive feature (removal drops accuracy 11.33 pp and Kappa from 0.6999 to 0.5237); that Pulse Pressure, when used as a replacement for the raw SBP/DBP pair, yields a 2.46 pp accuracy gain over the six-feature baseline; and that blood pressure features alone, without metabolic context, collapse to 57.14% - establishing Blood Sugar as non-substitutable and Pulse Pressure as a more efficient BP summary than either component in isolation. Our classifier outperforms prior published benchmarks: Ahmed (2020) at 75% and Okitikpi et al. (2025) at 73.5%.