What is it about?
This research presents CMSFusion, a Cross-Modal Spatial Fusion Network for adolescent mental health assessment. It aims to improve the assessment of depression, anxiety, and stress (D/A/S) by analyzing audio and visual behavioral features collected during conversational interactions. The framework combines audio and video information across multiple interview sessions using a Conformer-based architecture. It also introduces a Query-aligned Item Classifier (QIC) Head to capture fine-grained symptoms across the 21 items of the DASS-21 questionnaire, together with an Item-level Expert Ensemble (IEE) strategy to improve prediction accuracy. CMSFusion achieved 1st place in the DASS-21 (A2) track of the ACM Multimedia 2026 AdoDAS Challenge.
Featured Image
Photo by Veli Batuhan Aytaç on Unsplash
Why is it important?
Adolescent mental health screening is increasingly important, yet traditional assessments rely heavily on subjective questionnaires and clinical resources. What makes our work unique is its combination of privacy-preserving audio-visual features and fine-grained, item-level symptom modeling, enabling a more detailed assessment of depression, anxiety, and stress. By demonstrating the potential of multimodal AI for privacy-conscious mental health screening, this research may contribute to more objective, accessible, and scalable approaches to early identification and intervention.
Perspectives
For me, the most meaningful aspect of this work is exploring how multimodal AI can contribute to a better understanding of adolescent mental health while respecting individual privacy. This research has strengthened my belief that AI can serve as a valuable tool to complement traditional mental health assessments. I hope this work inspires further research toward more reliable, interpretable, and practical AI-based mental health screening systems.
Kaiwen Qin
Read the Original
This page is a summary of: CMSFusion: A Cross-Modal Spatial Fusion Network for Adolescent Mental Health Assessment, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3767308.3836622.
You can read the full text:
Contributors
The following have contributed to this page
