Integrating Gait With Multilevel Scale Representations for Depression Severity Prediction.

Depression is a severe mental health disorder with significant emotional and physical impacts. To identify and assess its severity, self-rating scales are commonly employed, which evaluate depressive symptoms through a series of questionnaire-based items. While these scales provide valuable subjective insights, gait analysis has emerged as a complementary, non-intrusive method for depression assessment, offering objective perspectives on an individual's mental state. Existing studies primarily use scale scores as ground truth while overlooking the semantic richness within the scale content. In this study, we propose a novel depression severity prediction framework that integrates gait features with scale content to enhance depression assessment. Specifically, the scale content is modeled as a multilevel semantic structure, comprising individual-independent and individual-specific components. Fusion strategies are tailored accordingly to align each type with the gait features. The individual-independent content, which provides general descriptions of depressive symptoms, is fused with gait features via a cross-attention mechanism to offer broad semantic guidance. In contrast, the individual-specific content, derived from participants' personalized responses, is used to align the fused features for more accurate and tailored prediction. We conduct extensive experiments on the D-Gait dataset, demonstrating that integrating scale content significantly enhances performance, with a notable 6.74% improvement in Concordance Correlation Coefficient compared to gait-only models.