A Cloud-Edge Collaborative Framework for EEG-Based Depression Recognition via Universal Pretraining and Hierarchical Quality Control

Resting-state electroencephalography (EEG) is attractive for depression recognition, but practical deployment still faces scarce labels, unstable signal quality, and heterogeneous resource budgets across end, edge, and cloud layers. We therefore propose a cloud-edge collaborative framework that integrates self-supervised pretraining, downstream fine-tuning, and a hierarchical three-stage quality-control ladder with coarse screening (Q1), statistical refinement (Q2), and geometry-aware refinement (Q3). The framework is evaluated by pretraining on a large public healthy resting-state EEG dataset from OpenNeuro (ds005385), main validation on selected data from the public TDBRAIN and MODMA multi-channel resting-state datasets, and controlled robustness testing on EEGdenoiseNet.Three findings emerge. First, on TDBRAIN, pretrained initialization clearly outperforms random initialization, reaching 0.8734 versus 0.6825. Second, hierarchical quality control strengthens weaker baselines, improving scratch accuracy from 0.6825 under No-QC to 0.7339 with Q1+Q2 and 0.7659 with Q1+Q2+Q3.Third, resource profiling supports layered deployment: Q1 runs at 6.78 ms per epoch, the classifier head has 2242 parameters, and the encoder and pretraining model reach 49,504 and 90,419 parameters, respectively. These results provide a quantitative basis and practical design principle for cloud-edge collaborative EEG depression recognition.