Short-form video platforms have become prominent spaces for mental health disclosure, peer support, and seeking information. They integrate text, visuals, and audio into a single stream of communication. However, much empirical research in digital mental health examines these modalities in isolation. This study introduces a pipeline combining automated multimodal feature extraction with Shapley value-based interpretability to analyze how text, visuals, and audio jointly influence engagement with mental health content. Applying this framework to 162,965 TikTok videos and 814,825 images about social anxiety disorder (SAD), we find that (1) visual facial expressions are more predictive of engagement than textual sentiment; (2) informational content attracts more attention than emotional support; and (3) cross-modal interactions exhibit threshold-dependent effects on reach. These findings advance our understanding of mental health communication in multimodal environments, demonstrating how platform algorithms privilege certain forms of mental health communication over others. Methodologically, this work contributes a reproducible and interpretable framework for multimodal research applicable across domains.

PDF