A Lightweight Multimodal Feature Alignment Framework for Depression Detection

Abstract

With the rapid advancement of artificial intelligence, computer vision, speech recognition, and natural language processing, automatic depression detection based on multimodal data has attracted increasing research attention. Compared with unimodal approaches, multimodal fusion leverages complementary information from speech, text, facial expressions, and other behavioral cues, thereby improving the accuracy and robustness of depression assessment. However, existing multimodal models are often computationally intensive and parameter-heavy, which limits their deployment on resource-constrained devices. In addition, semantic and distributional discrepancies across modalities pose significant challenges for effective feature alignment, adversely affecting fusion performance. To address these issues, this paper proposes a lightweight multimodal feature alignment framework for depression severity estimation. The proposed method constructs lightweight feature extraction networks and introduces a cross-modal feature alignment mechanism to enable effective mapping and fusion across heterogeneous feature spaces. While significantly reducing model size and computational complexity, the framework maintains competitive predictive performance. Experimental results on multiple public depression datasets demonstrate that the proposed approach achieves a mean absolute error (MAE) of 4.44 and a root mean square error (RMSE) of 5.77 in PHQ-8 score estimation, indicating strong generalization capability and practical deployment potential.