Building an auxiliary diagnostic and treatment efficacy prediction model for adolescent depression using machine...

Adolescent depression, characterized by high incidence rates and significant recurrence risks, has emerged as a major global public health concern. Current diagnosis relies predominantly on subjective questionnaires, lacking objective biomarkers and presenting challenges for early identification. Although exercise intervention is recommended as a safe and effective first-line non-pharmacological treatment, individual responses vary considerably. Approximately 30-40% of patients experience limited symptom improvement, necessitating reliable neurobiological markers to predict intervention response and enable precision exercise prescription. Resting-state electroencephalography (rs-EEG), with its non-invasive nature, low cost, and high temporal resolution, represents an ideal tool for identifying objective markers of depression. This study aims to construct a high-performance auxiliary diagnostic model for adolescent depression based on static and dynamic functional connectivity features of rs-EEG, combined with machine learning algorithms. It further seeks to identify neurophysiological biomarkers capable of predicting exercise intervention efficacy, thereby providing objective and practical technical support for early identification and individualized non-pharmacological interventions.

This study employed a machine learning framework combining diagnostic and predictive models. The diagnostic model aimed to identify depressed individuals among healthy adolescents (N = 155) and those with subclinical depression (N = 149), whilst the predictive model focused on the emotional response to a virtual reality-based exercise intervention within the same depressed cohort (N = 149), categorizing participants into high, moderate, and low responders. Six-minute resting-state EEG signals were collected, preprocessed, and segmented into 78 brain regions using the Schaefer template. These regions were then grouped into five core networks (sensory-motor, dorsal attention, salience, default mode, and central executive networks). Forty-five phase-locked functional connectivity features across the theta, alpha, and beta frequency bands were subsequently extracted. A three-stage feature selection method (random forest importance ranking, Pearson correlation redundancy removal, exhaustive search to determine optimal subsets) was employed to screen key neural biomarkers. Four classification models - logistic regression, support vector machines, random forest, and XGBoost - were constructed and compared. Ten-fold cross-validation and grid search were employed to optimize hyperparameters, with performance evaluated using accuracy, recall, precision, F1-score, and AUC-ROC. Finally, the SHAP framework was applied for model interpretability analysis, revealing the contributions and interactions of key brain network features.

Diagnostic model: The XGBoost model demonstrated exceptional performance in distinguishing SCD patients from healthy controls (AUC of 0.994, accuracy of 96.7%). Core biomarkers primarily concentrated on connections within the dorsal attention network (DAN), salience network (SN), and sensorimotor network (SMN) across the β and α frequency bands, with intra-DAN connections in the β band contributing most significantly. SHAP interaction dependency analysis further revealed complex nonlinear interaction patterns that may underlie, including a "fork-collider" synergistic structure centered on β-DAN and θ-SMN-DAN, cross-frequency inhibition of β-DAN by α-SN, and dynamic gating of β-DMN by α-SMN-DAN. The results of this study indicate that subclinical depression (ScD) manifests as a pattern of synergistic dysregulation involving multiple frequency bands and brain networks. In terms of predictive modeling, support vector machines (SVMs) performed best in predicting the response of ScD patients to exercise intervention, with baseline positive affect (PA) levels and functional connectivity between the somatosensory motor network (SMN) and the dorsal attentional network (DAN) in the β-band identified as key predictors. Furthermore, the patterns of feature interactions revealed by the model provide a testable framework for understanding the underlying model-derived neurophysiological patterns of subclinical depression from the perspective of multi-frequency, multi-network dyssynchrony; however, these results should still be interpreted as statistical associations rather than direct evidence of biological causality. SHAP analysis revealed an interpretable integration pathway ("behavior-driven → core coordination → dynamic gating → frequency band balance"), providing model-driven insights into the role of behavior-neural integration in personalized interventions.

This study systematically elucidates for the first time the intricate model-derived neurophysiological patterns of subclinical depression at the multi-frequency brain network level. It confirms the substantial potential of EEG functional connectivity as an objective biomarker while providing an interpretable neurophysiological basis for the precise selection of exercise interventions. These findings deepen our understanding of the early pathophysiology of depression and lay a robust foundation for developing low-cost, non-invasive tools for early screening and personalized treatment decision-making.