Capsule-augmented deep learning architectures for mental health detection from social media text

Mental health detection from social media text has attracted growing research attention due to the global rise in mental health concerns. Traditional deep learning models, such as Bidirectional Long Short-Term Memory (BiLSTM) networks and hybrid Convolutional BiLSTM (Conv-BiLSTM) architectures, have demonstrated strong performance in text classification tasks. However, these models often struggle to capture the hierarchical and spatial relationships that are intrinsic to linguistic data. To address this limitation, this study investigates the integration of capsule networks with BiLSTM and Conv-BiLSTM architectures for mental health detection. Leveraging a real-world Reddit corpus, we conduct extensive experiments comparing baseline BiLSTM and Conv-BiLSTM models with their capsule-enhanced counterparts. Furthermore, we explore the role of advanced loss functions, such as focal loss and contrastive loss, in addressing class imbalance and mitigating boundary blurring among semantically overlapping disorders. Our findings indicate that incorporating capsule layers significantly strengthens feature representation, leading to notable improvements in accuracy and F1-score across multiple mental health categories. The study focuses on six key disorders, including depression, anxiety, borderline personality disorder (BPD), and bipolar disorder. In addition, model interpretability is enhanced using Local Interpretable Model-agnostic Explanations (LIME), which highlights the critical linguistic features driving predictions, thereby improving transparency and reliability in mental health evaluations. • We propose a Conv-BiLSTM-Capsule architecture that effectively captures sequential, spatial, and hierarchical features for mental health detection from social media text. • To address severe class imbalance, we integrate SMOTE, contrastive loss, and focal loss, which improve learning for minority disorders. • The model is evaluated on a large Reddit mental health dataset covering six conditions, including depression, anxiety, bipolar disorder, autism, borderline personality disorder (BPD), and schizophrenia. • Results show notable improvements in minority-class performance (e.g., autism, bipolar) and better handling of boundary cases such as depression vs. anxiety. • We enhance interpretability by applying LIME explainable AI, highlighting key disorder-specific linguistic features that guide classification decisions.