Distill to Diagnose: Toward Stable and Reliable Cross-Modal KD Strategies for Depression

Multimodal fusion improves depression recognition but incurs high acquisition and deployment costs, limiting its practicality in real-world settings. Cross-modal knowledge distillation (KD) addresses this issue by using multimodal data during training while enabling unimodal inference at test time. However, existing studies mainly focus on architectural design and often overlook the stability and reliability of the transfer process under modality discrepancy and signal quality variation. In this work, we focus on stable and reliable cross-modal KD strategies for depression recognition. Specifically, we observe that large conditional variance in teacher features, conditioned on student-accessible modalities, can destabilize knowledge transfer. To address this issue, we introduce a conditional variance constraint to suppress such uncertainty and improve the stability of teacher-side supervision. In addition, we propose a quality-aware weighting (QAW) strategy to adaptively adjust the contribution of training samples according to modality reliability. These two strategies can be incorporated into the training process to provide more stable and reliable supervisory signals for the student model. Experiments on one synthetic dataset and two real-world multimodal depression datasets (EFVP and AADP) demonstrate consistent improvements in accuracy and F1-score, supporting more effective unimodal depression recognition in resource-constrained healthcare and everyday environments.