A Hybrid Transformer and CNN-Based Approach for Classifying Mental Health Disorders from Social Media Data

Abstract

Mental health disorders are a significant global concern, with increasing prevalence on social media platforms where individuals often share their experiences and emotions. This research presents a novel approach for classifying mental health disorders, specifically depression, anxiety, borderline personality disorder (BPD), and post-traumatic stress disorder (PTSD), using social media text. We propose a hybrid architecture that combines domain-specific transformer models, such as PsychBERT and MetaBERT, with Convolutional Neural Networks (CNNs) to enhance the model’s ability to understand mental health-related language and metaphorical expressions. The transformer models, pretrained on mental health and symbolic data, generate embeddings that capture the unique linguistic features in social media posts. These embeddings are processed through cascaded CNN layers to extract deep features, which are then concatenated and classified into mental illness categories. The model was evaluated using a balanced dataset comprising 40,000 social media posts, achieving an overall accuracy of 96% and an F1-score of 0.96. The proposed model outperforms existing state-of-the-art methods, including fine-tuned BERT and RoBERTa models, demonstrating superior performance in accurately classifying mental health disorders. The results highlight the effectiveness of leveraging domain-specific language models and CNNs for enhanced classification of mental health conditions in social media text. This study underscores the potential of advanced deep learning techniques in addressing mental health issues and facilitating early detection in real-world applications.