This experiment utilized a public dataset, namely the MODMA dataset (Mohammadi and Moradi, 2021), which was established by the Second Hospital of Lanzhou University. This dataset mainly consists of 55 participants, including a total of 26 outpatients diagnosed with depression (15 males and 11 females; aged 16–56 years), and 29 healthy controls (19 males and 10 females; aged 18–55 years). All MDD patients received a structured Mini-International Neuropsychiatric Interview (MINI) that met the diagnostic criteria for major depression of the Diagnostic and Statistical Manual of Mental Disorders (DSM) based on the DSM-IV. The dataset adopts a three-lead full-brain coverage EEG experimental protocol. According to the international 10–20 system electrode placement standard, three positioning points are selected on the forehead for electrode placement, with their specific pasting positions shown in the Figure 2. All subjects completed the Mini-Mental State Examination (MMSE) with the assistance of professional psychologists as a preliminary screening for depressive tendencies. If participants were at high risk of depression, they were required to additionally complete the Patient Health Questionnaire-9 (PHQ-9) to assess depression severity, while all basic information was collected. Candidate subjects were comprehensively determined to meet the experimental requirements based on self-rating scale data and inclusion criteria. Eligible subjects completed head cleaning under staff guidance and then wore detection equipment in a standard experimental environment. It should be specifically noted that: all subjects must not have taken any psychotropic drugs within 2 weeks before the experiment, and must not have other mental illnesses or organic brain injuries (such as epilepsy). Female subjects with depression must confirm that they are not pregnant. Meanwhile, the following conditions are excluded: lactating women, those taking contraceptives, individuals with a history of alcohol or psychotropic drug abuse/dependence within the past year, and individuals who have suffered abuse.

Sample Entropy (SE) is a method proposed by Richman (2011) in 2000 to measure the complexity of time series. According to the principle and formula of sample entropy, a higher entropy value of a time series indicates greater complexity; conversely, a lower entropy value implies higher autocorrelation of the time series. In recent years, entropy has emerged as a novel method for evaluating the complexity and irregularity of EEG signals in individuals with Depressive Disorder. Increased EEG signal entropy has been observed in patients with DDD, which indicates enhanced complexity and reduced predictability of brain activity (Chen et al., 2020; Čukić et al., 2020). The integration of this information-theoretic approach is regarded as a promising method for the assessment and monitoring of clinical depression (Murphy et al., 2020).

In the formula: r– similarity tolerance.

Variational Mode Decomposition (Dragomiretskiy and Zosso, 2013) is a novel adaptive signal decomposition technique. It is a non-recursive method that decomposes a multi-component signal into an ensemble of band-limited intrinsic mode functions (IMFs), also known as modes or components, with specific sparsity properties.

The key advantages of VMD include its ability to adaptively decompose non-stationary and non-linear signals, its robustness to noise and sampling, and its capability to handle different types of signals, including those with closely spaced frequency components. VMD has found successful applications in various fields, such as biomedical signal processing, fault diagnosis, and financial time series analysis. The main steps of VMD are as follows:

Use the alternating direction of the multiplier to calculate Equation 8 and continuously optimize by alternating and iteratively updating ukn+1, ωkn+1 ,λkn+1 to obtain the optimal solution of Equation 8. Among them,ukn+1 can be transformed into the frequency domain through Fourier transform, and we can get:

Where X is the constraint condition of û_k, u_k that is, ∑kuk=f; the purpose of the quadratic penalty factor α is mainly to reduce the signal interference of Gaussian noise;λ∧(ω) is the tolerance of the entire noise signal, which is mainly used to ensure that the signal after decomposition is not distorted;  f∧(ω) is the Fourier transform of f(ω),u^i(ω) is the Fourier transform of u_k(t).

Equation 9 can be transformed into the frequency domain through Fourier transform, and then the solution of u^kn+1(ω) can be obtained as:

In our proposed method, VMD is used to decompose the EEG signal into its intrinsic mode or component, which can then be used for component selection using the power spectrum, avoiding interference from redundant information.

Light Gradient Boosting Machine (LightGBM) is a highly efficient implementation of the gradient boosting decision tree (GBDT) algorithm, proposed by Ke et al. to address the limitations of traditional GBDT in handling large-scale datasets, such as high computational complexity and slow training speed [1]. Distinguished by two core optimization strategies–Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB)–LightGBM achieves significant improvements in training efficiency while maintaining or enhancing prediction accuracy, making it widely applied in fields like machine learning, data mining, and biomedical signal analysis (e.g., EEG-based depression detection).

The algorithm flowchart selected in this paper is shown in Figure 3. First, three-channel EEG signals are collected through the human prefrontal lobe. The collected signals undergo feature extraction via Variational Mode Decomposition (VMD), and appropriate feature components are calculated by combining the power spectrum. Finally, sample entropy is calculated for the obtained feature components. After obtaining the entropy features, they are fed into the LightBGM network for classification to ultimately determine whether the subject is in a depressive state or a depressive patient. The block diagram of its algorithm is shown in the Figure 4a.

As can be seen from Figure 4a, after inputting the original EEG signals, they are first decomposed using Variational Mode Decomposition (VMD) to obtain a total of 4 Intrinsic Mode Function (IMF) feature components. These four components undergo sample entropy feature extraction, and the resulting features are finally input into LightGBM for classification.

Considering the strong correlation between the prefrontal lobe and emotional processes, as well as mental illnesses, electroencephalogram (EEG) signals were collected via three electrodes. A common EEG acquisition device has three electrodes (Fp1, Fpz, and Fp2) on the prefrontal lobe. Data were recorded in a room free of loud noise and strong magnetism. Participants kept their eyes closed until their EEG signals were observed to be relatively stable, after which we began 90-s data acquisition, the sampling frequency is 250 Hz. In the processing of the MODMA dataset, the original hexadecimal data were first converted into decimal data. Then, the signals were filtered with a 1 Hz high-pass and 45 Hz low-pass finite impulse response (FIR) filter. Finally, the dataset used an adaptive noise canceller to eliminate eye-blink artifacts, thereby obtaining the noise-removed EEG signal data. The resulting EEG signals are shown in Figure 5, where HC represents healthy subjects and MDD represents major depressive disorder patients.

First, the data of each group were decomposed by VMD. The Variational Mode Decomposition (VMD) method is adopted mainly because the EEG activities in the δ, θ, α, and β frequency bands of patients with depression are generally higher than those of the normal control group. Moreover, the α and β frequency bands contain more depression-related EEG information than the low-frequency δ and θ bands. Therefore, decomposing the EEG signal into different frequency bands via VMD can effectively filter out interference from other frequency bands and ensure the validity of the signal. The VMD technique was used to decompose the non-stationary EMG signals into multiple frequency-band-limited IMFs, making each decomposed component easier to distinguish for emotion state classification. Taking the HC data as an example, parameter enumeration and optimization of the VMD algorithm were performed to obtain five groups of components. The decomposition results are shown in Figure 6. The characteristic components after VMD decomposition can more obviously reflect the variation trends of EEG signals at different frequencies, and the variation characteristics of EEG can be retained by extracting the effective fluctuation information of each modal component.

In this work, frequency analysis was conducted to determine the primary IMFs. Figure 7 shows the sample power diagrams of the primary IMFs. Since the frequency distribution of IMF5 differs significantly from that of the remaining components, only IMF1-IMF4 were selected as the primary IMF components for feature extraction. Additionally, it can be clearly observed from the power spectrograms that although the signal is decomposed into multiple groups of IMFs, highly discriminative information is retained only in a few IMFs. IMF2 and IMF3 exhibit higher energy density, while IMF1 and IMF4 have relatively lower energy content.

After obtaining the IMFs, we extracted sample entropy features from the main IMFs. This feature can effectively characterize the changes in EEG complexity and avoid Gaussian noise interference. After extracting the sample entropy features, since there are three channels in total, each channel is decomposed into 4 IMF channels, and sample entropy features are extracted, we finally obtained 12 feature vectors.

To ensure the accuracy of evaluation results, this paper uses Accuracy, Precision, Recall, F1-Score, Confusion Matrix, and average time consumption (T) as evaluation indicators for the lower limb gait phase recognition method. Here, TP represents the number of samples correctly predicted as positive by the model, FN represents the number of positive samples incorrectly predicted as negative by the model, FP represents the number of samples incorrectly predicted as positive by the model, and TN represents the number of samples incorrectly predicted as negative by the model.

The average recognition time evaluates the real-time performance of the model by calculating the recognition time of each sample.

To optimize the parameters of the LightGBM classifier model, an enumeration method was used to search for optimal parameter combinations of different numbers of leaves and learning rates, thereby determining the best values of hyperparameters. The influence of model hyperparameters on the classifier's accuracy is shown in Table 1. It can be seen from the table that the highest classification accuracy is achieved with the parameters of a maximum number of leaves of 50 and a learning rate of 0.01. Therefore, this set of hyperparameters will be used for subsequent model training and testing.

To verify the generalization ability of the model, the test set data without cross-validation was used for model inspection, and the model accuracy is shown in Table 2.

Combined with Table 2, it can be seen that after 5-fold cross-validation, the accuracy of the model adopted in this paper on each fold of the test set remains above 97.27%, with an average recognition accuracy of 97.42%. Among them, the fifth fold of data achieves the highest accuracy, while the third fold of data has relatively the lowest accuracy. The model exhibits good stability, and the accuracy on the test set can distinguish data of different categories, further verifying the effectiveness of the model. To fully exploit and utilize the temporal features in the training set and enhance the model's generalization ability, the trained model is applied to the test set for evaluation, and the confusion matrix shown in Figure 8 is obtained. The confusion matrix can intuitively reflect the classification performance of the model under various categories, where HC represents healthy subjects and MDD represents major depressive disorder patients.

As shown in Figure 8, the recognition accuracy of each phase in the test set is above 97.42%. During the learning process, the model needs to effectively extract and classify features from a large number of complex and interleaved features, but the overall recognition accuracy is high. Similar features in some data may lead to misjudgments. The experimental results in this paper show that the model can accurately determine the emotional state of the population, providing a universal judgment basis for related applications. Therefore, the effectiveness of the model using EEG signals for depression recognition is verified.

To further verify the effectiveness and superiority of the proposed method, this subsection conducts horizontal comparisons using algorithms that have achieved excellent results in emotion recognition and pattern recognition in recent years. The comparative methods cover both machine learning and deep learning approaches, all trained and tested on the same dataset. The comparison methods cover machine learning and deep learning methods, all of which are trained and tested on the same dataset. In terms of dataset division, we perform a sliding window on 90-s data samples with a length of 1 s, resulting in 4,950 samples. The dataset is segmented at 1-s intervals, mainly because there are many factors that ultimately affect the occurrence of MUS. A large proportion of these factors are caused by somatization and depression. This paper only discusses the technical means for preliminary screening of depression. With only one second of data, it can effectively determine whether an individual suffers from depression, thereby realizing the front-end triage of undifferentiated disorders, quickly screening out depressed patients from normal individuals, and improving the diagnostic efficiency. However, due to certain feature similarity between adjacent sliding windows, using the conventional 7:3 random division of samples in machine learning may easily lead to similarity between some samples in the test set and training set. Therefore, we divide the 4,950 samples according to the time dimension, with the first 50% of the samples as the training set and the last 50% as the test set. From the training set, 70% is extracted for training and 30% for validation. To ensure the effectiveness of the model, all input sample data undergo VMD component decomposition and sample entropy feature extraction. In this study, we used a laptop with an Intel i5-12400F @2.5 GHz CPU and an NVIDIA RTX 3060 GPU as the hardware environment. The software environment consists of Python with PyTorch 2.2.2. The algorithm comparison results are shown in Table 3.

For the CNN-LSTM model, in terms of network structure, we adopted a network structure where a 3-layer CNN network is connected in series with LSTM. Here, the “3 layers” refer to 3 modules, and each module includes a 3 × 3 convolution, regularization, a ReLU layer, and a global pooling layer, as shown in Figure 4b. We chose to use the Adam optimizer, set the learning rate to 0.001, set the total number of epochs to 100, and set the batch size to 64. The loss function used is the cross-entropy loss function. For the CNN-BiLSTM model, we only performed bidirectional processing on the LSTM model, and the remaining parameters are the same as those of the CNN-LSTM model. The experimental results show that the method proposed in this paper achieves the best performance in classification, with an average recognition accuracy of 97.42%, significantly superior to other comparative algorithms. In terms of model parameters, this method only has 0.32 M parameters, far lower than other deep learning algorithms, demonstrating its lightweight advantage.

In machine learning methods, relying on manually extracted features inevitably leads to partial loss of EEG features, directly limiting the model accuracy of traditional algorithms such as SVM and RF. However, the machine learning algorithm LightGBM, with its efficient decision tree mechanism, can not only achieve high classification accuracy but also maintain extremely small model parameters under the same EEG feature input, reflecting the effectiveness of lightweight models in feature utilization.

In the field of deep learning, 3-layerCNN-LSTM and 3-layerCNN-BILSTM models improve classification accuracy compared with traditional machine learning methods by fusing the temporal features of EEG signals, which verifies the critical impact of signal temporal information on classification performance. The training loss is shown in Figure 9. It can be seen from Figure 7 that Multi-Attention Convolutional Neural Network (MACNN) converges relatively slower than both CNN-LSTM and the proposed algorithm in this paper, and its final accuracy is also lower than the proposed algorithm. Meanwhile, in terms of time consumption, the MACNN algorithm takes longer. However, the proposed algorithm further breaks through the limitation of a single feature dimension through a multi-scale feature extraction strategy. In horizontal comparison, although the performance effect is slightly lower than that of 3CNN-Bilstm, it is stronger than 3CNN-Bilstm in terms of model lightweight and computational time. In summary, the proposed method shows significant advantages in three dimensions: classification accuracy, number of parameters and recognition time, and provides a more practical solution for the classification of depression based on EEG signals.

A one-way analysis of variance was used to measure the significant difference level between the comparative methods and the proposed method. As shown in Figure 10, there are significant differences (p ≤ 0.001) between the proposed method and other comparative methods, These comparison algorithms have lower model recognition accuracy than the proposed method on the same dataset, with only 3CNN-BiLSTM being slightly higher than the algorithm in this paper by 0.3%. However, as can be seen from Table 3, the computation time of 3CNN-BiLSTM is 2.9 ms, while that of the algorithm in this paper is only 2.23 ms. Considering both accuracy and timeliness, the algorithm in this paper has certain effectiveness in this classification task.

To further screen the influence of features and verify the effectiveness of each layer in the proposed model, we conducted ablation experiments, and the results are shown in Table 4.

In terms of dataset division, to avoid the situation where the dataset has high feature similarity due to sliding windows, we divided the entire sample into two parts in a ratio of 5:5. The first 50% of the total sample dataset is the training set, and the latter 50% is the test set. Meanwhile, we extracted 70% from the training set for model training, and 30% from the test set as the test set for the model. First, we used the original EEG data without any feature extraction and input it into the lightGBM model, achieving an accuracy of 93.78%. This is because there is a certain difference between patients and healthy people in binary classification data. Then, we added VMD for modal decomposition of the model, where K = 4. Each channel obtained 4 modal decomposition vectors, and then 12-dimensional features were input into the model for classification. Multiple measurements showed a classification accuracy of 95.27%. Finally, after adding the sample entropy, the overall accuracy increased by 2.15%, which proves the effectiveness of the sample entropy feature.

To further verify the effectiveness of the features, we selected three common features, namely the sample entropy feature, RMS, and PSD feature. After VMD decomposition into 12-dimensional vector data, we performed a sliding window with a window length of 20 data points, extracted feature vectors, and input them into the model for recognition. Tests showed that the sample entropy feature was slightly higher than the RMS and PSD features, thus proving the effectiveness of the model.