Multimodal Anxiety Discorder Detection Based on Clinical Interview

ACL ARR 2025 May Submission6496 Authors

Abstract: With the rapid development of artificial intelligence, multimodal methods have received increasing attention in the field of mental health disorder detection. Most of the existing research focuses on depression and schizophrenia, but there are relatively few studies on anxiety disorders. To further explore the clinical applicability of multimodal learning in anxiety disorder detection, we propose Multimodal Anxiety Detection via Clinical Interviews (MADCI), a framework designed to automatically identify anxiety disorders from real-world patient-doctor interview data. MADCI comprises three main components: modality-specific feature extractors, a hierarchical cross-modal attention fusion module, and a residual-enhanced multilayer perceptron classifier. In particular, the hierarchical cross-modal attention fusion module captures semantic correlations and complementary information across modalities by integrating cross-modal interactions at multiple levels, thereby enhancing the robustness and discriminative capacity of the fused representations. Experimental results on the MMDA anxiety disorder dataset demonstrate that MADCI achieves an accuracy of 87.13\% and an AUC of 88.36\%, significantly outperforming existing state-of-the-art multimodal baselines. These results underscore the effectiveness and clinical potential of our proposed approach in the automatic detection of anxiety disorders.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: multimodal applications; healthcare applications, clinical NLP;

Languages Studied: Chinese

Submission Number: 6496