Detection Methods for Depressive Disorders based on Resting-State EEG Signals

Abstract

Major depressive disorder (MDD) is primarily diagnosed through structured interviews, symptom-rating scales, and clinical judgment, whereas objective neurophysiological biomarkers have not yet been standardized for routine practice. Resting-state electroencephalography (rs-EEG) is noninvasive, relatively inexpensive, has high temporal resolution, and is suitable for repeated recordings. It has therefore become an important platform at the intersection of electronic information engineering and psychiatry. This review synthesizes representative studies of multicenter validation, functional brain networks, adaptive signal decomposition, convolutional neural networks, recurrent and spiking neural networks, Transformers, graph neural networks, and graph-convolutional Transformer architectures to analyze the complete technical chain of rs-EEG-based depression detection. The review first examines eyes-open and eyes-closed paradigms, sampling and referencing, filtering and artifact rejection, empirical mode decomposition (EMD) and variational mode decomposition (VMD), spectral and nonlinear complexity measures, functional connectivity, and graph-theoretic indices. It then compares support vector machines, random forests, convolutional neural networks, long short-term memory networks, spiking neural networks, self-attention, graph convolutional networks, and hybrid architectures. Representative findings are interpreted according to subject-independent partitioning, independent test sets, cross-site heterogeneity, data-leakage risk, and the comparability of evaluation metrics. Available evidence indicates that interchannel connectivity may reflect the network-level abnormalities of MDD more robustly than isolated channel power. The integration of graph-based spatial modeling with global temporal modeling is methodologically well motivated, but high accuracy in a small cohort should not be interpreted as clinical generalizability. Current work remains constrained by limited sample sizes, inconsistent diagnostic labels, distribution shifts across devices and sites, sensitivity to graph construction, insufficient interpretability, privacy barriers to data sharing, and the difficulty of lightweight deployment. Future studies should prioritize self-supervised pretraining, dynamic brain graphs, cross-domain and federated learning, uncertainty calibration, low-channel wearable systems, prospective clinical validation, and interpretable neurophysiological biomarkers. These directions are essential for moving rs-EEG-assisted diagnosis from algorithmic demonstrations toward reproducible, explainable, and deployable clinical decision-support tools.