A Depression Detection Method Based on Multi-Modal Feature Fusion Using Cross-Attention

Depression impacts about 3.8% of the global population, with over 75% untreated in low- and middle-income countries due to challenges in early diagnosis. This paper presents a new method for detecting depression using multi-modal feature fusion with Cross-Attention. Utilizing MacBERT for extracting lexical features from text and adding a Transformer module for task-specific context, this approach enhances model adaptability. Unlike simple concatenation of features, it employs Cross-Attention for integration, significantly improving detection accuracy and analysis of emotions and behaviors. A Multi-Modal Feature Fusion Network based on Cross-Attention (MFFNC) is developed, achieving 0.9495 accuracy on test data, a significant improvement. This method highlights technology's potential for early mental health intervention and its applicability to social media platforms for multi-modal processing tasks.