Multi-modal Depression Estimation using Adaptive Attentional Fusion Networks

Depression is one of the leading causes of disability worldwide, yet diagnosis still depends largely on subjective clinical interviews and self-report questionnaires. Automatic depression estimation using artificial intelligence provides a scalable and objective alternative. In this work, we explore a multi-modal framework based on a privately generated dataset, combining visual, acoustic, and textual features. An Adaptive ConvBiLSTM Fusion Network is proposed to capture complementary cues across modalities. The model employs CNNs and BiLSTMs for feature extraction, followed by hierarchical attention-based fusion layers to integrate multi-modal signals. Experimental results demonstrate that adaptive fusion outperforms unimodal baselines and standard late fusion approaches, providing more robust and interpretable depression estimation.