Cross-Model Attention and Multimodal Embedding Fusion for Depression Detection Using Clinical Interviews

Automatic depression screening from clinical interviews is a challenging task because depressive symptomscan appear across language, voice, facial behavior, and questionnaire responses. Existing systems often depend on a singlemodality or simple late fusion, which may miss complementary clinical evidence. This work proposes a cross-modelattention and multimodal embedding fusion framework for depression detection using clinical interview data. Theproposed framework represents each interview through text, audio, video-derived features, and PHQ-8-guided symptomevidence. The text branch uses contextual MPNet embeddings from interview responses, the audio branch capturesspeech-related cues such as prosody and vocal behavior, and the visual branch represents facial behavior usingexpression, gaze, head-pose, and action-unit features. These modality-specific representations are combined throughmultimodal embedding fusion, while cross-model attention assigns importance to complementary model outputs so thatthe final prediction is not dominated by one unreliable modality. The system is evaluated using depression-screeninglabels derived from PHQ-8 scores and clinical interview datasets such as DAIC-WOZ and extended DAIC-style data. Theofficial locked-test model achieved 76.6% accuracy, 0.593 F1-score, and 0.868 ROC-AUC on the DAIC-WOZ test split.The PHQ-guided out-of-fold fusion achieved 83.0% accuracy and 0.826 ROC-AUC, showing that symptom-aware crossmodel fusion can improve ranking and interpretability. The framework supports depression/non-depressionclassification, risk scoring, and modality-wise evidence analysis. Overall, the proposed approach improves screeningrobustness, interpretability, and controlled multimodal evaluation while remaining a research screening aid rather than amedical diagnostic tool.