Hierarchical Evidence-Gated Fusion With Cross-Task Consistency for Elder Depression Detection

This paper describes our submission framework for the elder track of the MPDD-AVG 2026 Grand Challenge, which focuses on multimodal personality-aware depression detection using audio-visual interviews, gait signals, and individual descriptions. Our approach organizes self-extracted speech and visual representations, official gait signals, and a compact structured profile vector into a hierarchical reliability-gated fusion framework. The neural path supports binary depression classification, ternary severity classification, and nine-item Patient Health Questionnaire (PHQ-9) score regression, while an auxiliary consistency verifier is introduced only for cross-task conflict cases. In addition, a severity-aware PHQ calibration strategy is applied at the final decision stage to enforce consistency between categorical severity and clinical score range. On the official evaluation results available to us, the complete system achieves a final score of 0.8208, with Binary Macro-F1 of 0.8516, Ternary Macro-F1 of 0.8839, and a raw PHQ concordance correlation coefficient (CCC) of 0.8389.