Uncertainty-Aware Evidential Multi-Modal Fusion Network for Large-Scale Multi-Center Major Depressive Disorder Classification

The clinical diagnosis of major depressive disorder (MDD) has long relied on subjective assessments, underscoring the need for objective and quantifiable automated diagnostic methods. Existing neuroimaging-based multimodal fusion methods face three core challenges: the lack of mechanisms that dynamically assess modality reliability, the absence of effective prediction uncertainty quantification, and site effects in multi-center data that constrain model generalizability. To mitigate these issues, we propose an uncertainty-aware evidential multimodal fusion network for large-scale multi-center MDD classification. We design deep functional feature extractors (DFFE) and dual-stream hierarchical feature fusion (DS-HFF) modules to encode functional magnetic resonance imaging (fMRI) and structural MRI (sMRI) features, respectively. An Evidential Multimodal Fusion (EMF) strategy based on Dempster-Shafer theory (DST) is designed to convert each modality's categorical support into subjective opinions with explicit uncertainty. Reliable cross-modal evidence integration is achieved through context-adaptive discounting and conflict-aware combination. Site-adversarial regularization is incorporated to learn site-invariant pathological feature representations. Evaluated on the REST-meta-MDD dataset comprising 1,601 subjects from 16 sites under a leave-one-site-out cross-validation (LOSO-CV) scheme, the proposed method outperforms recently published methods. Uncertainty analysis demonstrates superior calibration over Softmax and MC Dropout baselines, and the uncertainty-based rejection mechanism further improves the classification performance. Interpretability analysis identifies key brain regions highly consistent with established MDD pathophysiological findings, further validating the clinical reliability of the framework. Experimental results indicate that our proposed method has the potential to provide a robust, trustworthy, and interpretable solution for multi-center MDD diagnosis.