Multimodal Depression Detection Based on Modality Feature Decomposition and Contrastive Learning

Multimodal depression detection aims to identify depressive states by integrating complementary affective cues from audio, visual, and textual modalities. However, existing approaches often directly fuse multimodal features, which may introduce semantic interference, information redundancy, and insufficiently discriminative representations. To address these issues, this paper proposes a multimodal depression detection model based on modality feature decomposition and contrastive learning (MFDCL). Specifically, unimodal representations are decomposed into modality-shared and modality-specific features to distinguish cross-modal affective consistency from modality-dependent characteristics. Subsequently, sample-to-sample contrastive learning is applied within the shared feature space, where label-guided positive and negative sample pairs are constructed to enhance the discriminability between depressed and non-depressed states. In addition, a lightweight multi-scale attention fusion module is introduced to capture affective cues across different receptive fields. Experimental results on the DAIC-WOZ and E-DAIC datasets demonstrate the effectiveness of the proposed method. On DAIC-WOZ, MFDCL achieves Precision, Recall, F1-score, and Accuracy of 0.92, 0.90, 0.91, and 0.94, respectively, while on E-DAIC, it achieves 0.95, 0.85, 0.90, and 0.97. These results indicate that MFDCL effectively enhances feature discriminability and achieves competitive performance in multimodal depression detection.