The EEG data used in this study were collected from local hospital, and locally recruited individuals through professional screening. All participants in the experiment completed the Hamilton Depression Rating Scale - 17 (HAMD-17) before the formal data collection. All HC scored ≤7 on the HAMD-17, while patients with DD scored >17 on the same scale. All selected subjects were right-handed, with no other psychiatric disorders that could potentially impair brain function (such as dementia, schizophrenia, or anxiety disorders, except depressive disorder) and no physical impairments (such as severe cardiopulmonary, liver, kidney dysfunction, or autoimmune diseases). They had no history of drug or alcohol abuse and showed no signs of brain injury, with normal or corrected-to-normal vision. Each participant was prohibited from consuming alcohol and psychoactive drugs for 8 hours before the EEG recording. The EEG collection process occurred quietly without any other electromagnetic interference. The Ethics Committee of Zhejiang Normal University approved the study, and all participants provided written informed consent before participation.

Based on the above conditions, this study collected data from 70 patients with DD and 30 HC. The age range of the DD patients was 19 to 61 years, with an average age of (53.34±15.04). All subjects had been ill for more than one month. The age range of the HC group was between 21 and 57 years, with an average age of (37.70±13.32). There was no statistical difference in age between the DD group and the HC group, but the HAMD-17 scores showed significant differences. The overall research framework of this study is illustrated in Figure 1.

The EEG cap used in this study is Nicolet EEG TS215605. The electrodes are arranged according to the 10–20 international electrode system, retaining data from 16 EEG electrodes, as shown in Figure 1. It includes five brain interest regions: the frontal, temporal, parietal, occipital, and central regions. The experiment set the left and right earlobes (mastoid) as reference electrodes, with an EEG signal sampling frequency of 250Hz, ensuring that the impedance value of each electrode during the experiment is less than 5KΩ.

During EEG signal acquisition, it is very easy to be interfered with by various factors such as the environment, equipment, and physiology. These noises can negatively impact the signal quality. Preprocessing is a significant step in EEG data analysis, as it can effectively improve the signal-to-noise ratio, providing a reliable foundation for subsequent analysis and interpretation. The specific steps for preprocessing EEG in this study are as follows.

Feature extraction is the most common data analysis method in EEG experiments, specifically summarized as single-channel, dual-channel, and multi-channel EEG analysis methods. The single-channel EEG analysis method is the most common EEG analysis method, including time domain analysis, frequency domain analysis, and nonlinear dynamics analysis. Dual-channel EEG analysis methods are commonly used to analyze the correlation characteristics of two-lead EEG signals, mainly to explore the brain functional connectivity characteristics between different leads and between different brain regions, and the standard analysis methods include MI, PLI, and coherence analysis. The multi-channel EEG analysis method refers to analyzing the functional state of a specific brain functional region, mainly based on the brain functional network analysis of complex network theory, including clustering coefficient analysis, characteristic path length analysis, and small world attribute analysis.

Based on literature research, this study will adopt a multidimensional feature perspective, using three widely used EEG features, including PSD features, SE features, and PLI features, encompassing linear, nonlinear, and brain functional connectivity analyses. These features have been repeatedly proven effective in detecting mental disorders. However, no one has yet compared and analyzed the advantages and dis-advantages of EEG characteristics in the DD group from a multidimensional feature perspective. The following will provide a detailed introduction to the specific calculation processes of the three features.

To address the pronounced class imbalance inherent in neuropsychiatric datasets (elevated DD: HC ratio), this study implemented Cross-Cluster Replication (CCR), a hybrid framework synergizing energy-based minority neighborhood purification with structured oversampling, to optimize model robustness against distributional skew. CCR’s dual-phase architecture first implements density-aware data purification to eliminate noisy majority-class encroachments in minority sample vicinities, followed by cluster-informed synthetic instance generation that preserves neurophysiological feature covariance. This approach strategically balances recall-precision tradeoffs critical in clinical diagnostics, where false-negative minimization is paramount yet must not catastrophically compromise specificity. The algorithm’s hierarchical resampling mechanism ensures minority class representation aligns with the neurodynamic complexity of depressive phenotypes while maintaining electrophysiological plausibility constraints, thereby enhancing classifier generalizability beyond conventional undersampling/oversampling paradigms. The specific calculation processes of CCR are as follows:

First, for dataset X containing the majority class and the minority class, set the oversampling proportion N and the nearest neighbor parameter k.

Second, randomly select a sample X_i from the minority class samples and find the k nearest neighbors of this sample.

Further, for each minority class sample X_i, repeat the following steps N times (randomly select a nearest-neighbor sample X_nn from the k nearest-neighbors, generate a random number λ∈[0,1], and compute a new synthetic sample Xnew=Xi+λ×Xnn−Xi.

Finally, X_new is added to the new dataset and combined with the original dataset.

To strictly prevent data leakage and ensure the generalization capability of the models, we adopted a Leave-One-Out Cross-Validation (LOOCV) framework. In each iteration of LOOCV, one subject was held out as the test set, while the remaining N−1 subjects constituted the training set. This process was repeated N times (where N=100) so that each subject served as the test sample exactly once.

Building upon this validation framework, to ensure reproducibility and fair comparison across different classifiers, we implemented a consistent hyperparameter optimization protocol (46, 47). Considering the high computational cost of the LOOCV scheme and the risk of overfitting associated with exhaustive searches on limited data, we employed a Targeted Grid Search strategy. For each classifier, the optimization was constrained to a focused parameter space centered on established heuristic baselines. The optimal configuration for each LOOCV iteration was selected based on the peak validation accuracy within the training fold. The specific hyperparameter search spaces, final standardized configurations, and methodological justifications for all employed classifiers (including SVM, Random Forest, XGBoost, etc.) are explicitly detailed in Table 1. This comprehensive summary ensures that the experimental conditions are fully transparent and reproducible.

To quantify classification performance, we utilized the confusion matrix (Table 2), which categorizes predictions into True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN). TP and TN represent correctly classified positive and negative samples, respectively, while FP and FN denote misclassified instances. While Accuracy (calculated as shown in Equation 12) serves as the primary metric, we expanded the evaluation to include Recall, Precision, F1-score, and the Area Under the ROC Curve (AUC) to provide a robust assessment (calculated as shown in Equations 13–15). Given the LOOCV design where each test fold contains a single sample, these metrics were derived using a prediction pooling strategy, in which predictions from all subjects were aggregated to compute global performance measures. Furthermore, to quantify statistical uncertainty, the Accuracy is reported with its 95% Confidence Interval (Accuracy (%) [95% CI]), calculated based on the normal approximation of the binomial distribution.

This study addresses the challenge of high-dimensional neurophysiological feature analysis (n = 608 features encompassing PS, SE, and PLI), with 64 features derived from PSD (16 channels × 4 bands), 64 from SE (16 channels × 4 bands), and 480 from PLI (120 channel pairs × 4 bands), through a rigorous SVM-RFE framework optimized for differentiating DD (n = 70) from HC (n = 30). The methodology iteratively eliminates redundant features via backward selection, prioritizing features based on maximum margin weight magnitudes. To mitigate stochastic instability inherent in single-iteration rankings, the algorithm implements 100 iterations for the 100 subjects, generating probabilistic ranking matrices (100 × n), with final feature prioritization determined by modal frequency analysis across rank positions (The method for determining the final ranking results is to select the most frequently occurring feature from the first column of the 100 × n feature ranking results as the first feature in the final feature ranking and to select the two most common features from the first and second columns as the second feature, following this rule to determine the final feature ranking results). Optimal feature subset identification follows an incremental forward inclusion protocol, systematically evaluating classification performance with sequentially expanded feature combinations.

While emerging techniques such as graph neural networks and attention-based models have demonstrated strong capabilities in modeling complex and high-order feature interactions, their effective deployment in EEG-based psychiatric research demonstrably requires substantially larger datasets to mitigate overfitting and is often accompanied by significant computational burden and reduced interpretability (48–50). In contrast, the SVM-RFE framework adopted in this study provides a computationally efficient and highly interpretable feature-selection strategy that aligns well with the characteristics of moderate-sized clinical EEG datasets. Its linear kernel and explicit feature-ranking mechanism enable transparent identification of neurophysiologically plausible biomarkers, a requirement of particular importance in depression-related research where clinical interpretability and reproducibility remain central objectives. Furthermore, this approach offers a well-calibrated balance between dimensionality reduction and the retention of critical neurodynamic information, which is especially relevant given the markedly higher dimensionality of PLI connectivity features relative to PSD and SE. By distilling the multidimensional EEG space into a compact and clinically meaningful subset, the SVM-RFE framework ensures both analytical rigor and neurophysiological plausibility, thereby serving as a methodologically sound choice for the present investigation.

This study employed a leave-one-out evaluation framework and multiple machine learning classifiers (SVM, CatBoost, XGBoost, LightGBM, RF, and KNN) to distinguish DD patients from HC. The classification accuracy of each individual feature type and their combination is shown in Figure 2, with the top-performing classifiers summarized in Table 3. XGBoost yielded the highest accuracy for PSD (80.33±29.77%), whereas SVM yielded the highest accuracy for SE (63.82±26.30%) and PLI (89.00±21.37%). When all three feature types were combined, SVM achieved an accuracy of 89.95±20.44%.

To provide a more comprehensive assessment of classifier performance, additional metrics—including Recall, Precision, F1-score, AUC, and Accuracy (%) with 95% Confidence Intervals—were calculated by aggregating predictions across all leave-one-out folds. As presented in Table 4, PLI exhibited the strongest single-feature performance (Recall 91.36%, Precision 93.01%, F1-score 92.18%, AUC 95.32%, Accuracy 89.00% [95% CI: 88.03–89.97%]). The combined-feature model further improved performance, achieving a Recall of 93.99%, Precision of 93.26%, F1-score of 93.63%, AUC of 96.49%, and the highest Accuracy at 90.90% (95% CI: 90.01–91.79%).

To evaluate the impact of the data augmentation strategy on handling class imbalance, this study conducted a comparative analysis between the proposed CCR method and the non-CCR dataset. As detailed in Table 5, the inclusion of CCR consistently enhanced classification performance across feature categories. Specifically, for the optimal combined feature set, accuracy improved from 85.01±30.37% (non-CCR) to 89.95%±20.44% (CCR), and PLI accuracy increased from 84.89±29.25% to 89.00±21.37%. These results confirm that CCR effectively mitigates majority class bias and is essential for the model’s robustness.

As shown in Figure 3, classification accuracy increased steadily from 4-s to 10-s windows, with the 10-s window yielding the highest performance under the SVM classifier. A subsequent decline was observed at the 12-s window.

Given SVM’s excellent performance, we adopt the RFE feature selection method. Specifically, its feature ordering is computed for each subject. One hundred locally ranked features are obtained among 100 subjects, and the 100 locally ranked features are linearly summed to obtain the globally ranked feature sequence. Using the SVM classifier, features are added one by one for the global optimal feature sequence, and the results obtained are shown in Figure 4. Following feature selection, an optimal feature subset comprising 153 features achieved a classification accuracy of 94.48%. The detailed classification results for each feature within the optimal subset are presented in Table 6. Among the selected features, the subset included 17 PSD features, 10 SE features, and 126 PLI features. Notably, there were 11, 6, and 45 features corresponding to the beta frequency band, representing the largest proportion of features within each category.

To enhance interpretability, we visualized the discriminative PLI connections identified by the model across the four frequency bands (Figure 5). Red lines represent hyper-synchronization in DD relative to HC, whereas blue lines indicate hypo-synchronization. Theta and alpha1 bands showed stronger fronto-parietal and fronto-temporal synchronization in DD, while alpha2 and beta bands exhibited reduced connectivity primarily within centro-parietal and fronto-central regions. These patterns align with the distribution of high-ranking PLI features identified through SVM-RFE.

This study uses the optimal feature subset and SVM machine learning method to identify DD, resulting in the following scatter plot, as shown in Figure 6. The recognition accuracy for each individual ranged mainly between 80-100%, indicating stable performance. Excitingly, the number of subjects with an accuracy of over 80% was 92 (the total is 100), and the number of subjects with an accuracy of over 50% was 99 (the total is 100), indicating that the model has good predictability and stability.

This study indicates that a 10-second fixed-length window provides an optimal balance between achieving stable estimates of nonlinear complexity and connectivity features while preserving the quasi-stationary characteristics of EEG signals, thereby yielding robust classification performance in depression diagnosis. The choice of a fixed-window strategy—as opposed to adaptive or sliding-window approaches—was deliberate and grounded in several methodological and practical considerations. Historically, a wide range of time-window lengths has been applied in EEG signal processing (45, 51, 52), with fixed-length windows and sliding-mode windows, being among the most commonly used strategies (53). Fixed windows offer computational simplicity, standardized temporal resolution, and stable statistical conditions for estimating spectral and phase-based connectivity measures, which are essential for cross-subject comparability and the extraction of reproducible biomarkers in depression-related EEG research (54). They also minimize the risk of feature instability induced by variable overlap or step sizes in sliding windows. Sliding-mode windows introduce temporal overlap to improve continuity and sensitivity to gradual signal changes but may increase feature redundancy and attenuate low-frequency components (55). Although fixed windows may seem methodologically simpler than adaptive approaches, they are particularly well aligned with the objectives of depression-related EEG analysis, where standardized temporal resolution is crucial for cross-study comparability and the extraction of robust biomarkers. In the context of depression—where neural alterations often manifest as sustained oscillatory and network-level dysregulations—the controlled, parameter-invariant nature of fixed windows enables a clear assessment of how window duration per se influences feature discriminability. The controlled parameterization afforded by fixed windows enables a precise examination of how window duration influences multidimensional feature stability and classification performance, thereby establishing a reliable baseline for subsequent methodological innovation.

Substantial evidence from emotion recognition studies also indicates that affective neural states typically unfold over time spans exceeding 10 seconds (56–58), and similar considerations apply in automated EEG recognition of epileptic activity (59). From a computational perspective, longer windows improve efficiency by reducing segmentation frequency, whereas shorter windows better capture transient neural dynamics but impose higher computational demands.

Prior work consistently shows that window length critically shapes classification outcomes, underscoring the need for disorder-specific temporal optimization. For example, Luo et al. reported peak performance using a 10-second window for anxiety-related connectivity analysis (60), while Fang et al. identified 12-second windows as optimal for mixed anxiety–depression classification (61). However, their study did not delineate condition-specific optimal windows for DD and anxiety subgroups. Given the distinct biological, psychopathological, and genetic mechanisms underlying these disorders (1, 62, 63), generalized temporal parameterization across diagnostic categories remains methodologically unsound, necessitating disorder-specific optimization. Within the comparative framework of this study, the 10-second window demonstrated consistently robust classification performance across the tested resolutions, supporting its use as a practical reference scale for DD identification. This exploration of temporal parameterization not only contributes to the discussion on diagnostic precision but also provides novel methodological insights for temporal feature engineering in psychiatric EEG research. The observed variation in performance across window sizes highlights the conceptual importance of this parameter and warrants further investigation with larger cohorts to establish its definitive optimal value.

This study establishes that multidimensional integration of EEG features significantly enhances DD classification accuracy compared to single-dimensional approaches. By synergistically combining these complementary neurophysiological dimensions, our support vector machine (SVM) model achieved 89.95% accuracy, outperforming single-dimension benchmarks. This performance gain reflects the capacity of multidimensional features to capture distinct yet interrelated pathological mechanisms: PSD quantifies localized oscillatory abnormalities in cortical excitability, SE characterizes neural adaptability through signal complexity, and PLI maps interregional network dysregulation. Previous studies relying on isolated features, such as Li et al. (64) (PSD alone) and Avots et al. (65) (linear/nonlinear feature combinations), demonstrate the inherent limitations of single-dimensional analyses in addressing DD’s neurobiological complexity. Crucially, our findings align with emerging evidence that DD manifests as a multidimensional disorder requiring concurrent evaluation of rhythmopathology, complexity collapse, and network desynchronization (66).

Notably, PLI emerged as the most discriminative feature type (89.90% accuracy), consistent with prior findings that PLI-based connectivity dominates optimal biomarker sets in DD (67) and reflects aberrant phase synchronization patterns reported in clinical cohorts (68). The spatial distribution of key PLI connections in our results (particularly the pronounced frontal–cingulate and frontal–parietal interactions) further supports the notion that DD involves disrupted executive–limbic network coordination, aligning with evidence of impaired top-down regulation in depressive neuropathology (66). While PLI alone demonstrates strong diagnostic value, our multidimensional fusion framework achieved even higher performance (94.48%), corroborating studies showing that integrating PSD, SE, and PLI captures complementary aspects of oscillatory alteration, complexity reduction, and network-level dysregulation (40).

Although this study used multidimensional features to construct the model, the overall accuracy was still not high before feature selection. We implemented SVM-RFE to eliminate low-weight features, with repeated leave-one-out cross-validation to identify the optimal feature subset, the optimal classification accuracy of 94.48% was achieved with 153 features. Notably, several studies have reported remarkably high classification accuracies of 99.08% (69) and 98.76% (70), which may be related to the different dataset partitioning methods used, such as subject crossover data division, which has a risk of data leakage, leading to an overestimation of the model’s actual performance. In contrast, the leave-one-out method can reduce such risks (71). In addition, feature selection and optimization are also commonly seen in the application of machine learning in other fields. There are many methods for feature selection, which tend to improve machine learning performance (72, 73). In the domain of DD diagnosis, Hierarchical Clustering and Spectral Network Fusion method significantly outperformed conventional feature selection approaches in classifying resting-state EEG signals for DD (74). Consistent with these findings, SVM-RFE method has been utilized to predict early treatment responses in DD, observing enhanced sensitivity compared to single-level prediction models (75). In summary, a review of the application of machine learning in the diagnosis and efficacy prediction of mental disorders (43, 76, 77), feature optimization is essential for improving the accuracy of machine learning. It can enhance model performance and increase classification accuracy. More efficient and reliable models can be constructed by employing reasonable feature selection. Our experimental results further validate that systematic feature selection constitutes an indispensable strategy for building efficient and reliable computational models.

Building on the optimized feature space, SVM-RFE revealed a feature subset dominated by Beta-band (13–30 Hz) activity, accounting for 64.7% of PSD, 60% of SE, and 35.7% of PLI features. This frequency-specific distribution suggests that abnormal Beta oscillations constitute a key electrophysiological signature of depressive disorders. This pattern is also consistent with findings from prior studies and advanced computational models (e.g., temporal–frequency attention), which frequently identify high-frequency hypersynchronization as an important discriminator in psychiatric conditions (48–50). The prominence of these Beta features points to specific pathophysiological mechanisms: the dominance of Beta-PSD reflects excessive local cortical excitability, while the prevalence of Beta-SE indicates rigid, low-complexity neural information processing. Collectively, these findings support the view that Beta-mediated cortical–limbic dysregulation constitutes a core electrophysiological signature of depressive pathology.

The limitation of this method is that our current dataset comprises 100 subjects, including 70 with DD and 30 HC, and the sample size remains limited. Future studies should expand the dataset to enhance reliability for developing an effective DD diagnostic method. In addition, EEG data were acquired using a 16-channel clinical montage, which may limit spatial resolution compared with high-density (32–64 channel) systems. Although this configuration reflects real-world clinical practice and has been shown to capture stable large-scale spectral and connectivity patterns, future work should further validate the proposed framework using higher-density EEG recordings.