An uncertainty-aware ensemble of support vector machine and Gaussian process models for maternal mental health risk prediction using psychosocial and clinical risk factors

Abstract Postpartum depression (PPD) is a major maternal mental health condition that can adversely affect mothers and infants if not identified and managed early. This study developed and evaluated an uncertainty-aware ensemble framework integrating Support Vector Machine (SVM) and Gaussian Process (GP) models for maternal mental health risk prediction using psychosocial, demographic, socioeconomic, obstetric and clinical factors. A pooled dataset comprising 14,008 retrospective maternal records from Kawempe National Referral Hospital and Mukono General Hospital in Uganda and the publicly available PERI_DEP dataset was analysed. Preprocessing included missing-value imputation, categorical encoding, feature engineering, Mutual Information-based feature selection, Kernel Principal Component Analysis and Synthetic Minority Oversampling Technique-based class balancing. All preprocessing and oversampling procedures were restricted to the training data. The Hybrid SVM–GP framework was compared with Logistic Regression, Random Forest and CatBoost using stratified five-fold cross-validation and a held-out internal test set. To assess potential target leakage, analyses were conducted using both the complete feature set and a leakage-free feature set that excluded individual Edinburgh Postnatal Depression Scale items and directly related depression-symptom variables. In the full-feature analysis, Random Forest achieved the strongest discrimination, with an accuracy of 97.75% and a Receiver Operating Characteristic Area Under the Curve of 0.996. The Hybrid SVM–GP achieved an accuracy of 93.54%, a Receiver Operating Characteristic Area Under the Curve of 0.982 and the lowest Expected Calibration Error (ECE) of 0.015, whereas Random Forest achieved lower Brier score and log loss. The Hybrid SVM-GP also provided explicit case-level uncertainty information based on the normalised predictive entropy of the fused probability, with GP posterior variance providing a complementary measure of uncertainty in the latent prediction. These findings indicate that the Hybrid SVM–GP may complement models such as Random Forest by identifying potentially ambiguous predictions requiring further assessment, rather than providing universal performance superiority. SHapley Additive exPlanations from the leakage-free analysis identified family relationships, socioeconomic conditions, social support and reproductive history as influential contributors to model predictions. However, the retrospective design and absence of independent external validation limit immediate clinical application. Prospective external validation and clinical-utility assessment are required before routine maternal healthcare deployment.