CROSS-ATTENTIVE MULTIMODAL TRANSFORMER ARCHITECTURE FOR MENTAL HEALTH DISORDER DETECTION FROM TEXT

Mental health disorders such as depression, anxiety, and stress increasingly manifest through language used in digital communication.Leveraging Natural Language Processing (NLP) for automated detection offers a scalable alternative to conventional clinical assessments.This paper proposes a Cross-Attentive Multimodal Transformer Architecture that integrates BlueBERT for semantic representation, EmotionBERT enhanced with BiLSTM-CNN for emotional modeling, and VADER for sentiment polarity estimation.A cross-attention mechanism aligns emotional and sentiment-derived features with clinical semantics to improve contextual understanding.To address class imbalance, the dataset is reformulated into a balanced distribution using stratified sampling and targeted augmentation.Training employs Focal Loss with class weights, AdamW optimization, and mixed-precision acceleration.Experimental evaluation on the balanced dataset achieves 92% accuracy and a 0.92 macro F1-score, demonstrating substantial improvement over unimodal baselines and imbalanced training settings..The results highlight the effectiveness of attention-based multimodal fusion for reliable mental health disorder detection.