An Explainable Hybrid Model for Decoding Silent Mental Health Symptoms through Social Media Interaction and Textual Withdrawal Patterns

Mental health (MH) disorders, particularly depression and social withdrawal, represent critical global challenges, undermining both individual well-being and societal productivity. Social media provide a unique lens to capture digital behaviours that may serve as early indicators of MH states. While conventional diagnostic methods are often subjective and resource-intensive, digital behaviour analysis offers scalable, non-invasive alternatives. Yet, existing studies frequently isolate emotional, relational, and temporal dimensions, limiting predictive accuracy and interpretability. This study proposes digital behavioural continuum theory (DBCT), a framework integrating these dimensions to model MH states holistically. A hybrid machine learning architecture is developed, combining graph neural networks (GNNs) for relational structures, recurrent neural networks (RNNs) for temporal sequences, and Valence Aware Dictionary and Sentiment Reasoner (VADER) for the sentiment analysis technique to extract affective signals from user-generated content. Model transparency is ensured through Shapley additive explanations (SHAP), enabling identification of the most influential behavioural markers. Results demonstrate that emotional features (e.g., sentiment scores, sad reaction ratios) exert the greatest predictive influence, followed by temporal signals such as posting frequency and response latency, while relational attributes contextualise social withdrawal. The proposed model achieves an F1-score of 90.4%, a precision of 89.7%, and a recall of 91.2%, significantly surpassing baseline approaches. Importantly, the datasets analysed were not clinically diagnosed but were curated to reflect real-world social media behaviours associated with potential mental health signals. By advancing an interpretable, data-driven framework, this research bridges theoretical innovation with practical application, enhancing digital MH monitoring and supporting early, scalable interventions.