Dynamic Brain Network Characteristics of Adolescent Depression Symptoms: Analysis and Identification from Graph Theory Metrics to Spatiotemporal Attention Networks

Background Depressive symptoms in adolescents represent a major challenge in global public health, with rising prevalence rates significantly impacting academic performance, interpersonal relationships, and long-term mental health development. Existing neuroimaging studies indicate that depression involves dynamic dysregulation across large-scale brain networks, particularly abnormalities in networks related to emotional regulation, self-referential processing, and cognitive control. However, traditional static functional connectivity analysis struggles to capture millisecond-level dynamic brain reorganization processes. Meanwhile, EEG-based dynamic functional connectivity research remains relatively limited in adolescent populations, lacking systematic multi-method, multi-frequency comparisons and the development of AI-assisted diagnostic models. These gaps hinder in-depth understanding of the neurodynamic mechanisms underlying depressive symptoms and the creation of objective early identification tools. This study aims to systematically reveal the dynamic brain network abnormalities in adolescents with depressive symptoms through multi-frequency dynamic functional connectivity analysis, graph theory metrics, and deep learning models. It explores frequency-specific and method-dependent characteristics while constructing a high-precision, objective EEG-based intelligent recognition model. This research provides new evidence for elucidating the neural mechanisms of depressive symptoms and advancing clinical diagnostic support. Methods A total of 131 adolescents with depressive symptoms (ScD) and 155 healthy controls (HC) were recruited, and 6-minute resting-state EEG data were collected. Three methods—phase locking value (PLV), phase coherence (COH), and phase lag index (PLI)—were employed to construct dynamic functional connectivity matrices across δ, θ, α, and β bands. Modularity, intra-/inter-network connectivity strength and its time-varying characteristics, along with brain state dynamics metrics, were extracted for intergroup statistical comparisons. Concurrently, a Spatio-Temporal Dynamic Attention Network (ST-DAN) model integrating multi-frequency attention mechanisms, 1D-CNN, and bidirectional LSTM was designed. Classification performance was evaluated using 5-fold cross-validation. Results Graph theory analysis reveals distinct frequency band gradients in brain network abnormalities: α band abnormalities are minimal and method-dependent (COH shows occipital lobe enhancement, while PLI indicates reduced intensity with increased fluctuations). ; δ and β band abnormalities are confined to enhanced internal connectivity within the occipital network (PON) and abnormal frontal-occipital interactions, accompanied by state-dependent rigidity; θ band abnormalities are most widespread, encompassing modular temporal instability, enhanced internal integration in posterior brain regions (temporal TEN, occipital PON), disrupted cross-network coupling between anterior-posterior and motor-posterior regions, and reduced state flexibility. Results from the three methods were highly complementary, consistently highlighting the central role of PON and θ oscillations. Within the ST-DAN model, PLV demonstrated optimal performance (80.5% accuracy, AUC 0.875); θ monoband outperformed multi-band fusion (86.2% accuracy), suggesting risks of attentional bias and noisy shortcut learning. Conclusion This study integrates multi-method dynamic analysis with deep learning to reveal significant frequency-specific abnormalities between adolescent depressive symptoms and brain networks, particularly highlighting the biomarker role of theta band and occipital network abnormalities. This provides new insights into the neurodynamic mechanisms of adolescent depression and lays a theoretical foundation for developing high-precision EEG-based intelligent diagnostic tools. Furthermore, this study systematically evaluated the efficacy of functional connectivity metrics (COH, PLV, PLI) in SCD identification using the ST-DAN model. Results indicate that PLV outperforms other metrics in balancing accuracy, sensitivity, and specificity. Additionally, multi-frequency band fusion models demonstrate inferior performance compared to θ-band-only models, revealing potential risks of noise-driven shortcut learning. These findings not only validate the core physiological significance of the theta band in cognitive dysfunction but also provide optimized feature selection strategies for EEG-based early screening.