SAGE: Stability-Aware Graph-Based Ensemble Feature Selection for Explainable Postpartum Depression Risk Prediction

Personalized explanations matter for clinical use: The paper demonstrates that traditional machine learning models for PPD prediction often provide only opaque global explanations that aren't useful for individual patients. SAGE addresses this by using LIME-based explanations to provide personalized risk assessments for each woman, making the AI system more actionable for clinicians.

Psychological and socioeconomic factors are stronger PPD predictors than demographics: The research found that factors like depression screening scores (EPDS and PHQ-9), feelings about motherhood, and abuse history were the main predictors of postpartum depression, while demographic characteristics had less influence. This suggests interventions should focus on these modifiable psychological and social factors rather than demographic profiling.

Feature selection stability is crucial for trustworthy AI in healthcare: The paper emphasizes that unstable feature selection—where different features are chosen each time the model runs—undermines clinical trust. SAGE uses "bootstrap stability weighting" to ensure the same important features are consistently identified, which is essential for doctors to rely on the system's recommendations.

Responsible AI can work with imbalanced real-world data: PPD prediction involves class imbalance (fewer cases of depression than non-depression), which typically degrades AI performance. SAGE achieved strong results (87.96% accuracy, 0.88 AUC) while handling this imbalance through GAN-based oversampling, demonstrating that responsible design can address practical healthcare challenges.

Fewer features can mean better generalization: The system achieved high performance using only 16 features instead of larger feature sets, suggesting that interpretable, streamlined models may generalize better across different populations—important for deploying PPD screening tools in low- and middle-income countries where the burden is highest.