Abstract
<title>Abstract</title> <p>Automatic depression estimation is frequently approached through multimodal fusion of audio, text, and visual cues. However, audio and text streams are often noisy, unavailable, or undesirable in practice owing to environmental interference, transcription errors, and privacy constraints. Therefore, this study aims to develop and evaluate an integrative, video-based assistive tool that models the non-uniform, multi-scale temporal characteristics of depression-related facial dynamics across diverse clinical interviews. We present the micro-macro temporal alignment network (MMANet), a multi-cue visual deep learning framework for depression severity estimation. It incorporates a dual-scale unified selector to identify informative temporal windows at both micro and macro scales from interview recordings. The resulting representations are aligned through supervised contrastive learning to promote cross-scale consistency while preserving complementary cues. We evaluate MMANet on three diverse clinical interview benchmarks—DAIC-WOZ, E-DAIC, and CMDC—to assess its cross-cultural robustness, using mean absolute error (MAE) and root mean square error (RMSE) as primary metrics. The final analysis includes 264 unique participants. Using exclusively visual features, MMANet achieves an MAE of 3.83 and an RMSE of 5.29 on the DAIC-WOZ evaluation set; 4.15 and 5.33 for E-DAIC, and 3.04 and 4.06 for CMDC. Component analyses demonstrate that importance-aware temporal selection consistently outperforms random and uniform sampling approaches, and joint micro-macro modeling with cross-scale alignment yields the lowest prediction errors. These findings suggest that scalable, video-based integrative modeling of facial dynamics could serve as a privacy-preserving assistive tool for depression screening, though prospective validation in real-world clinical settings is required before implementation.</p>