A Predictive Model for Pregnancy Loss Among Reproductive Aged Women

Abstract

Pregnancy loss remains a critical global public health challenge, affecting approximately 23 million women annually and accounting for 15.3% of all clinically recognized pregnancies worldwide, with the highest burden concentrated in sub-Saharan Africa and other low-resource settings. Despite its prevalence, clinically deployable, explainable predictive tools capable of integrating multiple risk domains at the point of individual care remain largely absent from routine antenatal practice. This study developed, implemented, and evaluated PregLoss PredictorAI, a machine learning-based clinical decision support system for pregnancy loss risk stratification among reproductive-aged women. A dataset of 1,187 patients with eleven clinical and demographic features was used to train and evaluate three supervised ensemble classifiers (XGBoost, LightGBM, and Random Forest) under 5-fold stratified cross-validation. The XGBoost model was selected as the best-performing classifier, achieving a weighted AUC-ROC of 99.95%, accuracy of 98.90%, weighted F1-score of 0.9890, High-risk sensitivity of 98.52%, and High-risk specificity of 99.16%. SHAP analysis identified BMI, pre-existing diabetes, mental health concern, heart rate, and blood sugar as the five most influential predictors, findings consistent with established clinical evidence. The trained model was deployed as a 3-tier web-based application comprising a Python FastAPI backend, React.js clinical interface, and PostgreSQL database. The system provides patient-level SHAP explainability alongside risk classification, bridging the persistent translational gap between machine learning research and clinical practice. These findings demonstrate that a multi-domain, explainable, low-cost predictive system for pregnancy loss risk is achievable and clinically viable.