Privacy-preserving and explainable multimodal detection of depression and anxiety

Privacy exposure and limited interpretability remain major barriers to multimodal screening of depression and anxiety. PP-XMDA integrates federated optimization, differential privacy, adaptive attention fusion, uncertainty estimation, and multi-level attribution within a unified framework. The experimental implementation evaluates the three modalities available across the benchmark corpora: text, audio, and visual information; physiological and behavioral inputs are retained only as architectural extension pathways. Across MMDA, MMPsy, and DAIC-WOZ, the strongest depression-detection result reaches F1 $$=0.806$$ and AUC-ROC $$=0.873.$$ Under the selected privacy configuration of $$\epsilon=4.0,$$ the membership-inference attack success rate decreases from $$0.542$$ for federated learning without differential privacy to $$0.352,$$ while F1 decreases from $$0.806$$ to $$0.794$$ (1.49% relative loss). Missing-modality tests, cross-dataset transfer (F1 $$\geq 0.742$$ ), subgroup auditing, and perturbation-based explanation metrics indicate that privacy protection and predictive utility can be jointly optimized. Attention weights are reported as modality-contribution indicators rather than causal explanations, and the outputs are positioned as model-auditing and decision-support information rather than validated clinical diagnoses.