Early Detection of Postpartum Depression Using Explainable Deep Learning Models: A Comparative Study of DNN, GATE, and SAINT Architectures

Postpartum depression (PPD) affects up to 52.3% of women in Nigeria and sub-Saharan Africa, yet remains critically underdiagnosed due to cultural stigma, resource constraints, and the limitations of traditional screening tools such as the Edinburgh Postnatal Depression Scale. This study developed and comparatively analyzed three deep learning architectures—Deep Neural Network (DNN), Gated Attention Network (GATE), and Self-Attention Network for Tabular Data (SAINT)—to provide an explainable and deployable screening solution. Using a dataset of 1,503 postpartum records comprising sociodemographic and emotional variables, the models were trained using the Adam optimizer with dropout regularization. The DNN emerged as the superior architecture, achieving high performance across all evaluation metrics: 96% accuracy, 96% F1-score, and an ROC–AUC of 0.980, outperforming both GATE (93%) and SAINT (89%). To address the “black-box” barrier in clinical AI adoption, SHAP (SHapley Additive exPlanations) analysis was integrated, revealing that age (51.3% predictive contribution) and emotional indicators (48.7%) were the primary risk determinants. This level of transparency enables clinicians to interpret and justify individual risk forecasts. The final DNN model was implemented as a Flask-based web application, providing a real-time, scalable screening tool suitable for low-resource environments. The findings demonstrate that explainable deep learning can effectively address the PPD screening gap in sub-Saharan Africa. Future work should prioritize external validation with local Nigerian datasets and the integration of multi-class severity grading to further refine clinical decision support.