X²-Gait: A Framework for Dual-Level Interpretable Gait-Based Subthreshold Depression Recognition

Early subthreshold depression identification is critical for intervention but hampered by subjective assessments. Gait analysis offers a promising objective biomarker, but existing computational methods struggle to integrate multi-scale motor patterns and lack the interpretability required for clinical trust. This paper introduces X²-Gait, a novel deep learning framework designed to address these challenges. X²-Gait pioneers a hierarchical, context-aware mechanism that extracts features at multiple granularities, from fine-grained joint kinematics to holistic body posture. Crucially, it designates macro-level postural features as a 'contextual anchor' to dynamically guide the interpretation of micro-level movement features via a gating network. This architecture uniquely enables dual-level interpretability, providing both a high-level postural assessment and specific, corroborating biomechanical evidence for each prediction. Evaluated on a dataset of individuals with subthreshold depression and healthy controls, X²-Gait not only achieves state-of-theart classification accuracy (70.26%) but also generates clinically meaningful, hierarchical explanations. This work represents a significant step toward developing transparent, trustworthy, and non-intrusive tools for mental health screening.