Explainable Machine Learning Framework for Early Prediction of Adverse Pregnancy Outcomes in Low-and Middle-Income Countries

Adverse pregnancy outcomes, including pre-eclampsia, preterm birth, low birth weight and stillbirth, remain a leading source of maternal and perinatal death, and roughly nine in ten of these deaths occur in low- and middle-income countries (LMICs) where specialist obstetric assessment and laboratory testing are scarce. Risk scoring that works at the first antenatal contact, using only information a frontline health worker can collect, could redirect scarce referral capacity toward the pregnancies most likely to deteriorate. A transparent reproducible framework was built and interrogated for early prediction of a composite adverse pregnancy outcome on a simulated LMIC antenatal cohort of 10,800 pregnancies (10.6% event prevalence) whose datagenerating process was calibrated to published epidemiological parameters and restricted to twenty-five variables obtainable at a first booking visit. Seven classifiers were tuned with randomised search under stratified cross-validation. Interpretation combined a model-agnostic Monte-Carlo Shapley estimator, permutation importance and partial-dependence analysis. Discrimination, precision-recall behaviour, calibration, the Matthews correlation coefficient and the Brier score were reported, and models were compared with DeLong, McNemar and Wilcoxon tests. Gradient boosting gave the strongest discrimination (ROC-AUC 0.792, 95% CI 0.761-0.819; PR-AUC 0.416), reaching 78.3% sensitivity at 64.1% specificity at a screening-oriented operating point; DeLong tests confirmed its advantage over every model except the neural network. Shapley analysis identified chronic hypertension, antenatal-visit intensity, prior pregnancy loss, blood pressure, proteinuria and haemoglobin as the dominant drivers, matching obstetric expectation. The framework runs on low-cost, first-visit inputs, produces a per-patient explanation a midwife can act on, and maps predictions to a three-tier referral pathway, making it a plausible decision-support aid for primary antenatal care in resource-constrained settings pending prospective validation.

Keywords: Machine Learning (ML), Pregnancy Outcomes, Explainable AI, Maternal Health, LMICs, Risk Prediction.