As pressures in modern society rise, mental health disorders are becoming more prevalent. However, the shortage of psychiatrists has strained medical resources, increasing doctors’ workload and impacting care quality. Early depression detection and reducing doctors’ burden are thus critical challenges. This study uses the DAIC-WOZ dataset, which includes audio, facial/body features, text transcripts, and PHQ-8 scores for 189 interviews, to explore multimodal mental health assessment. By applying HuBERT and RoBERTa pre-trained Transformer models in a dual-channel attention architecture, the model achieved 0.75 accuracy, highlighting the potential of deep learning for efficient, accurate early interventions in clinical practice.