Hybrid Deep Learning Approaches for Social Media-Based Mental Health Detection

Social media platforms like Reddit are good indicators for the detection of mental health disorders since the users of these platforms express the ideas and emotions by posting what they think and feel. Identification of stress and depression from such platforms is necessary for early detection and support. A hybrid deep learning approach is introduced by enhancing transformer models such as BERT and Mental BERT by adding extra- linguistic features. Emotional tone from NRC Emotion Lexicon, psychological understanding from LIWC, and topic model representations from LDA and Top2Vec are added with BERT embeddings through the use of a multimodal adaptation gate. This improves the model's capability to recognize the nuances in mental health-related posts. For model stability, label smoothing is introduced to prevent overconfidence and improved calibration. The methodology is compared with real-time datasets collected directly from social media platforms such as Twitter and Reddit. Experimental results validate that the use of extra-linguistic features greatly enhances the classification performance with accuracy gain, precision, and recall. This step not only improves stress and depression detection but also provides deeper understanding of the language features of mental health-related posts. By combining transformer models and adding language features and enhancing calibration, the methodology ensures consistent and stable predictions, and this methodology is thus an extremely stable tool for real-time detection and analysis of mental health.