DEEP TRANSFORMER ARCHITECTURES FOR EARLY DETECTION OF MENTAL HEALTH DISORDERS FROM MULTIMODAL DATA

Abstract

Mental health disorders have emerged as one of the most pressing global health challenges, significantly affecting individual well-being and societal productivity. Traditional diagnostic approaches are often subjective, time-consuming, and lack scalability, highlighting the need for automated, data-driven solutions. Recent computational methods, including machine learning and deep learning, have shown potential, but most remain limited to unimodal analysis or suffer from high complexity and poor generalization. To address these challenges, this study proposes a deep transformer-based multimodal architecture for the early detection of mental health disorders. The framework integrates text, speech, and physiological signals through modality-specific encoders (BERT, CNN-BiGRU, 1D-CNN) and employs a transformer-based fusion mechanism to capture intra- and inter-modal dependencies. Experimental evaluation on a benchmark multimodal dataset demonstrates that the proposed model achieves superior performance compared to traditional and unimodal baselines, with an overall accuracy of 88%, a macro-F1 score of 86.8%, and an AUC-ROC of 0.91. These results confirm the effectiveness of transformer-based multimodal fusion in improving early diagnosis of mental health conditions. This work presents a scalable, interpretable, and efficient framework that supports large-scale mental health screening and enables timely intervention by clinicians.