BOTTLENECKCROSS-ATTENTIONMAMBA NETWORKFORMULTIMODALDEPRESSION DETECTION

Abstract

Depression detection using artificial intelligence has gained significant attention due to the limitations of traditional clinical assessment methods. This project explores the use of Mamba-based State Space Models for efficient and scalable multimodal depression detection. We implement and analyse BCM-Net, a progressive fusion ar chitecture designed to model long-range temporal dependencies in audio-visual data while maintaining low computational complexity. By evaluating baseline models and studying their performance on large-scale benchmark datasets, this work aims to iden tify research opportunities for improving accuracy, efficiency, and fusion strategies. Akey objective is to develop lightweight and accessible depression detection models, enabling practical deployment in real-world, resource-constrained settings.