A Symptom-Aware Framework for Postpartum Depression Risk Detection Across Heterogeneous Textual Data

Abstract

Postpartum depression (PPD) is a significant maternal mental health condition that can adversely affect maternal well-being, mother–infant bonding, and infant development, highlighting the importance of early risk detection. Although natural language processing (NLP) enables automated analysis of depression-related textual data, many approaches rely on single-domain learning and provide limited evaluation across heterogeneous textual environments. This study proposes the Clinical-Linguistic Symptom-Aware Postpartum Depression framework (CLiSA-PPD), integrating contextual representations from DistilRoBERTa with clinically guided linguistic features capturing anxiety, sadness, sleep disturbance, maternal distress, social isolation, and self-harm risk. An attention-guided fusion mechanism combines contextual and symptom-aware representations for PPD risk classification. The framework is evaluated using a PPD questionnaire dataset converted into narrative-style text and an external depression-related Reddit dataset. Comparative evaluation against Logistic Regression, Linear Support Vector Machine, Bidirectional Long Short-Term Memory, and DistilRoBERTa demonstrates competitive transformer-level performance. CLiSA-PPD achieves an accuracy of 0.9304, recall of 0.9859, F1-score of 0.9567, Matthews Correlation Coefficient of 0.7896, and Brier Score of 0.0562, indicating strong sensitivity and low probabilistic prediction error. Cross-domain evaluation reveals sensitivity to domain and class-distribution shifts, highlighting limitations in discrimination across heterogeneous textual environments. Ablation, explainability, calibration, statistical significance, and robustness analyses assess the contribution, reliability, and interpretability of the proposed framework.