Multi-Granularity Graph Convolution Network for Major Depressive Disorder Recognition

Major depressive disorder (MDD) is the most common psychological disease. To improve the recognition accuracy of MDD, more and more machine learning methods have been proposed to mining EEG features, i.e. typical brain functional patterns and recognition methods that are closely related to depression using resting EEG signals. Most existing methods typically utilize threshold methods to filter weak connections in the brain functional connectivity network (BFCN) and construct quantitative statistical features of brain function to measure the BFCN. However, these thresholds may excessively remove weak connections with functional relevance, which is not conducive to discovering potential hidden patterns in weak connections. In addition, statistical features cannot describe the topological structure features and information network propagation patterns of the brain’s different functional regions. To solve these problems, we propose a novel MDD recognition method based on a multi-granularity graph convolution network (MGGCN). On the one hand, this method applies multiple sets of different thresholds to build a multi-granularity functional neural network, which can remove noise while fully retaining valuable weak connections. On the other hand, this method utilizes graph neural network to learn the topological structure features and brain saliency patterns of changes between brain functional regions on the multi-granularity functional neural network. Experimental results on the benchmark datasets validate the superior performance and time complexity of MGGCN. The analysis shows that as the granularity increases, the connectivity defects in the right frontal(RF) and right temporal (RT) regions, left temporal(LT) and left posterior(LP) regions increase. The brain functional connections in these regions can serve as potential biomarkers for MDD recognition.

Recently, some studies [22], [23], [24] utilize deep neural networks to explore the relationship between the EEG channel location and depression, e.g.Seal et al. propose a deep convolution neural network framework called deep learning-based convolution neural network (CNN) (DeprNet) for detecting depression using EEG [25].To understand which channels are related to depression, DeprNet applies different spatial coordinate systems.Although DeprNet explore the relationship between spatial channels and MDD from the perspective of coordinate system, it did not explore the relationship between spatial location and brain maps of MDD.To learn the relationship between MDD brain map and spatial location area, Liu et al. propose a signal preprocessing method by converting EEG signals into brain maps, and propose a low complexity network architecture (one layer CNN and one layer GRU) to recognize MDD and normal control (NC) [26].Similarly, Goswami et al. also pay attention to the relationship between the time, frequency and space of EEG signals with depression by converting EEG activity into a series of topologically preserving multi-spectral images, and propose a method to recognize MDD from multi-channel EEG [27].However, [26] and [27] both ignore some long-distance brain connections in the steps of spatial location.Simply defining neighbors according to distance may lead to the loss of important information, as well as the importance of information transmission between brain regions and brain salient patterns for MDD recognition.Brain functional connection can simulate the dynamic interaction between brain neurons, so it can effectively avoid the above problems.
To learn the information transmission in the brain space and the brain core topological structure, some researchers attempt to utilize the graph neural network (GNN) to learn the brain topology and brain functional connectivity network (BFCN) for recognizing MDD, i.e.Zhu et al. utilize correlation to construct the BFCN by extracting the four linear features of EEG, namely activity, mobility complexity and power spectral density, apply the attention mechanism to calculate the brain connection in BFCN and input it into the GNN to recognize MDD [28].However, there are a large number of weak connections and pseudo connections in the BFCN, which often blur the topological structure of core connections.Thus, the threshold-based binary method to solve this problem may be one of the best candidates for removing weak connections [29].Zhang et al. construct BFCN based on phase lag index (PLI), transform BFCN into a binary brain network (BBN) using threshold with a large difference in the proportional threshold, and extract static statistical features from the BBN, such as the average shortest path length, clustering coefficient, small world index, etc.However, this threshold method may excessively remove weak connections with functional relevance, which is not conducive to finding hidden patterns in weak connections.In addition, static statistical features can not describe the topological structure of the brain network, brain salient patterns and information transmission between brain regions.
To solve the above problems, a MDD recognition method based on multi-granularity graph convolution network (MGGCN) is proposed.The main contributions of this paper are as follows: 1) Compared with the existing method that recognizes MDD based on static statistical features of BFCN, there is a defect in ignoring the brain core topology and the significant pattern of information transmission between brain regions.This paper proposes a multi-granularity method to continuously learn the brain network topology, and combines the brain network core topology with the GNN to recognize MDD for the first time, achieving a relatively advanced recognition effect.2) Compared to existing methods that roughly filter weak and pseudo connections in BFCN, there is a drawback of roughly filtering out weak connections with functional correlation.This paper proposes a MDD recognition method based on a multi-granularity graph convolution network, which achieves the effect of preserving the weak connections with functional correlation while removing the noise.The organizational structure of this paper is as follows.Section II introduces the MDD recognition framework based on MGGCN; Section III introduces the relevant information of the datasets applied in this paper, as well as the specific steps and results of the experiment; Section IV is a brief summary of research work.

A. Overview of MGGCN Model
To learn the core topological structure of the brain network and the brain salience pattern in depth, remove the noise in the BFCN while preserving the weak connection with functional correlation, a MDD recognition based on MGGCN is proposed.
As shown in Fig. 1, we preprocess the original resting state EEG signal; Secondly, the wavelet packet transform (WPT) has a high time-frequency resolution, so as to utilize WPT to decode and reconstruct the EEG signal into Delta, Theta, Alpha, Beta, and Gamma frequency bands, and retain the original undivided All frequency bands; Phase synchronization (PS) is more sensitive to phase change than other brain network construction methods, there by, apply phase locked value (PLV) and PLI to construct the BFCN; However, there are a lot of redundant and pseudo connections in the constructed BFCN.To learn the overall coordination of the brain and the core brain network topology, multi-granularity can both remove redundant connections and prevent the removal of core connections.Thus, transform BFCN into sub BBNs pyramid with different granularity; Train the sub BBNs pyramid of the original All band of MDD and NC constructed by the two methods to the proposed MGGCN.And then test the core topology recognition under different human states by learning the models of the core topology differences of different subjects' brain networks; Finally, softmax activation function is adopted to recognize MDD and NC.
The core part of MGGCN model is shown in Fig. 2 and describe in detail below:

B. EEG Frequency Processing
In the preprocessing stage, considering that the length of the original EEG signal of all datasets and the experiment are easily affected by the previous stimulus, the EEG signals of the original MDD and NC are normalized, and the EEG signals of the two are intercepted, respectively.The unified experimental signal is designed for the experimental paradigm with a duration of 80 seconds, and the EEG signal is processed in segments using a time window of 10 seconds.We apply FIR

TABLE I WAVELET COEFFICIENT INDEX AND FREQUENCY BANDS
band-pass filter 0.5 Hz∼40 Hz to filter EEG signals and utilize REST re-reference, and FastICA to remove ocular artifacts.
WPT is utilized to convert the original EEG signal into Delta, Theta, Alpha, Beta and Gamma frequency bands.We utilize the Wavelet Toolbox in MATLAB for wavelet transform.Daubechies4 (db4) is adopted as mother wavelet to decompose EEG data.The most important point is that db4 is closer to the EEG generated by normal brain activity.Therefore, the db4 is applied to decompose the EEG signal into six layers of wavelet, and then the wavelet coefficients per 1Hz are extracted in the sixth layer.
Input EEG signal S, and the WT can be expressed as vector matrix: among them, it contains N wavelet transform coefficients, and W is an orthogonal matrix composed of basis vectors.
among them, α 0 represents the wavelet coefficient of layer 0, m represents the number of resolution layers, and the sampling interval is related to the number of resolution layers.The sampling interval of layer m is equal to 2 m .G represents the low pass filter, H represents the high pass filter, and d represents the wavelet coefficients obtained after each layer decomposition.
C. Computational Multi-Granularity BBN Pyramid 1) Construction BFCN: Secondly, the BFCNs based on PLV and PLI are constructed respectively.
The research of BFCN is an important research topic in the field of brain computer interface (BCI).The BFCN regards the EEG electrode as a node in the network, and constructs the relationship between electrodes through different methods as the weighted connection of the BFCN [30], [31].Functional connectivity simulates the dynamic coordination and time interaction between different neurons in the brain [32], [33], [34].
Candra et al. found that the segmentation of 3-12 seconds can better describe emotions in their study on the classification of emotions based on the size of EEG segmentation [35].
Comprehensive analysis shows that a 10 second segmentation is a good choice.Haartsen et al. investigated the performance of and debiased weighted PLI (dwPLI) in measuring the reliability of EEG alpha connections with different time periods and numbers.They found that whole brain dwPLI has higher reliability in many shorter time periods (1 or 2s), while whole brain PLI has higher reliability in fewer longer time periods (5 or 6s) [36].
Similarly, Velde et al. also studied the reliability of PLI and dwPLI in BFCN [37].They found that compared to global PLI, global dwPLI had higher inter subject variability and lower reliability.Based on the above viewpoints, we chose a longer period of 10 seconds to construct BFCN using PLI with high reliability.
In addition, Ghazani and Phan proposed a graph convolutional neural network for analyzing EEG signals, and evaluated various connection methods such as Corr, Icoh, dPLI, IPLV, PLI, wPLI and dwPLI [38].They found that PLV outperformed other methods in estimating connectivity, while PLI method outperformed wPLI and dwPLI.Therefore, we chose PLV and PLI to construct BFCN to explore the differences in the core topology of the brain network between MDD and NC.
Each EEG signal is regarded as a node in the BFCN, and PLI and PLV are respectively utilized to calculate the edges between nodes in the BFCN.
Assuming that the EEG signals of any two electrodes are x(t) and y(t), the phases of these two signals at instant t are expressed as φ x (t) and φ y (t), respectively, and the phase change of the instant t is expressed: among φ x (t) and φ y (t) is the expansion of EEG signals x and y under time series t.
PLV is calculated as: PLI is a method to measure the asymmetry of phase difference distribution between two signals [39].
The asymmetry index of phase difference distribution can be obtained from the phase difference (t) of t time series.
2) Transformation Multi-Granularity BBN Pyramid: So far, we have obtained the BFCN based on PLV and PLI in all bands.There are a lot of weak connections and pseudo

D. Detailed Structure of MGGCN
To overcome above disadvantages, MGGCN is designed to remove weak connections and pseudo connections in BFCN, at the same time, it prevents filtering out the core connections and affecting the extraction of brain significant patterns.

Algorithm 1 MGGCN Model
Input: the weight of BFCN W i , the filter F w , the bias b and label La Output: Ŷt and Loss end while 18: end for The architecture of temporal convolution network (TCN) [41], [42] is proved to be particularly effective in predicting events with time components, such as the development of critical diseases and sequential operation of single electronic medical records [43], [44].TCN has some advantages over other neural networks.Firstly, it is faster than the LSTM based recurrent neural network model; Secondly, TCN is better at capturing temporal dependencies, and because of the use of convolution, it can capture local information; Last but not least, the size of receptive field can be flexibly adjusted.Different granularity sequences are utilized to learn the brain network core topology, therefore, the updating sequence of different granularity is applied to design the MGGCN, and the four layers MGGCN are utilized to deeply learn the core topology of the brain network: As shown in Fig. 3, the single MGGCN model is composed of the Dilated Causal Convolution, ReLU, Layer Norm and Spatial Dropout modules, and each block is composed of L convolution layer sequences.The result of dilated convolution at granularity t is: where each layer has the same number of filters F w , which enables us to combine activations from different layers using skip connections later.F w size is equal to the number of channels in the experimental dataset.Each layer consists a set of dilated convolutions with rate parameter s, a nonlinear activation f (.) and a residual connection than combines the layer's input and the convolution signal.The filters are parameterized by W (i) ∈ R F w ×F w and bias vector b ∈ R F w .The output of each block applies a set of skip connections to sum Z (0) ∈ R F w ×T .It is concluded that: where L is the number of convolutional layers.
After the sub BBN pyramid or the MGGCN of the previous state is input into the model, the dilated causal convolution is performed.The number of hidden layers is N N N ], and the expansion coefficient d is 2 i (i=1,2,4, . . ., n, n ≤ N ).The number of input channels N is the number of EEG signal channels, the number of output channels after convolution is N * N .Optimizer select the Adam optimizer.
To understand the role of multi granularity in overall model MDD recognition, we conducted ablation experiments in the experimental section and conducted experiments on the benchmark dataset.

E. MDD Recognition
The pooling layer adopts maximum pooling, which reduces the size of spatial information and improves operation efficiency.Reducing spatial information can also decrease the number of parameters and the risk of overflow.Additionally, it helps to achieve invariance of translation-rotation scale.
The collection of filters in each layer are defined as W = W (i) Fl i=1 for W (i) ∈ R d×F l−1 with a corresponding bias vector b ∈ R Fl .Given the signal from the previous layer E (l−1) , we compute activations E(l) with Finally, in the Dense layer, all core topology of the brain networks are and the Softmax activation function is utilized recognize MDD and NC.
where weight matrix U ∈ R C×F w and bias c ∈ R C .Cross_entropy is selected as the loss function of the model.
where Ŷt is the predictive value of the model, La is the label, and W is all the parameters of the model, α is the coefficient of the regular term.The cross_entropy function cr oss_entr opy( Ŷt , La) aims to measure the difference between the actual label and the predicted value of the model, while the regularization term α∥W ∥ 2 aims to reduce the over fitting probability of the model's learning parameters.

III. EXPERIMENTS AND ANALYSIS A. Materials
The public dataset MODMA is utilized to construct BFCNs to recognize MDD.24 MDD (female/male=11/13, 30.88 ± 10.37 years old) are the experimental subjects, and 29 NC (female/male=9/20, 31.45 ± 9.15 years old) are the control group.There is no significant difference in age (t=0.214,p=0.832) and gender (x2=1.224,p=0.269) between the two groups.The experimental data are recorded continuously by hydrocel geodesic sensor network (HCGSN) with 128 channels of Cz reference at a sampling frequency of 250 Hz [45].
In dataset MPHCE, 64 subjects are recruited, including 34 MDD and 30 NC.The status of subjects is confirmed by talking with professional doctors, including men/women=38/26, 39.4 ± 14.11 years old.EEG equipment adopts 19 electrode caps, sampling frequency is 256Hz, and EEG data are collected in the closed eyes resting state, open eyes resting state and task state.After screening the dataset, the status of all data sets is consistent, the resting EEG data of 25 NC and 23 MDD are utilized [46].
In the dataset EDR, 121 subjects are recruited, including 46 MDD and 77 NC.The diagnosis method is doctor's consultation and Beck depression scale judgment, in which male/female=47/74, 18.9 ± 1.2 years old.The EEG equipment utilizes 66 lead EEG electrode caps with a sampling frequency of 500 Hz.The resting EEG data is also applied to remove some subjects with missing data and mild depression patients.40 MDD and 67 NC are selected for MDD recognition experiment [47].
All subjects were diagnosed by doctors' dialogue, Beck Depression Inventory and other medical evaluation scales.To investigate the serious impact of depression on the brain, subjects with mild and moderate depression were excluded and subjects with major depression selected as MDD subjects.

B. Comparison Results on Benchmark Datasets
To demonstrate the impact of MGGCN in MDD recognition, we compare its performance with advanced methods using Accuracy (Acc), Precision (Prec), Recall, and F1_Score on benchmark datasets including MODMA, MPHCE, and EDR.
To provide the MGGCN with more comprehensive functional connectivity network information and facilitate visualization of BFCN, the MGGCN incorporates the complete BFCN as the feature.
We divide the dataset into 70% for model training and 30% for testing purposes.Furthermore, we perform 5-fold cross-validation.Table III shows our comparison results with traditional machine learning methods based on BFCN feature extraction and advanced GICN methods.Our MGGCN model achieves better recognition performance and lower time complexity than other methods due to the multi-granularity capture of the core topology of the brain network and the flexible receptive field of the model.
1) Result on MODMA Dataset: On the MODMA dataset, we compare the MDD recognition performance of PLI+SVM [12], PLI+PLV+SVM [13], and correlation+GICN [28] with that of MGGCN proposed by us.The analysis shows that the features fused by two BFCN construction methods in PLI+PLV+SVM [13] are better than those extracted by PLI+SVM [12] from traditional BFCN static statistical features, which supplement some complementary information so as to the MDD recognition effect is better.Correlaton+GICN [28] extracts BFCN features and utilize GCN method to recognize MDD.Compared with traditional machine learning method based PLI+SVM [12] and PLI+PLV+SVM [13], it has a better MDD recognition effect by deeply learning the relationship between brain network features and MDD.However, they all ignore the effect of redundant connections on feature extraction, as well as the role of brain core network structure and brain salient patterns.Therefore, the MDD recognition effect of the above work is lower than our MGGCN model.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE III COMPARISON OF MDD RECOGNITION EFFECT (%) AND TIME COST(S)
2) Result on MPHCE Dataset: On the MPHCE dataset, the analysis shows that the performance comparison results of PLI+SVM [12] and PLI+PLV+SVM [13] are consistent with the result of MODMA dataset, but the MDD recognition effect of Correlation+GICN [28] is weaker than that of PLI+PLV+SVM [13].The reason may be that the number of EEG channels in MPHCE dataset is 19, and SVM has high robustness in small data samples.However, our MGGCN still has some advantages, because our model has flexible receptive fields, which can adjust the input range of causal dilation convolution and can flexibly learn the core topology of the brain network and brain salient patterns with different data sizes.
3) Result on EDR Dataset: On the EDR dataset, the analysis shows that the comparison results of PLI+SVM [12] and MDD recognition of other works are consistent.However, in terms of Acc and Recall, the MDD recognition effect of Correlation+GICN [28] is stronger than that of PLI+PLV+SVM [13].In terms of Prec and F1_Score, the MDD recognition effect of Correlation+GICN [28] is weaker than that of PLI+PLV+SVM [13].This may be because the EEG channel number of EDR dataset is 64 and Correlation+GICN [28] is not sensitive to false negative.Our MGGCN still has some advantages, which shows that our MGGCN model is relatively stable, and the flexible local receptive field can ensure the MDD recognition effect of the model.
It can be seen from the synthesis that our MGGCN model is superior to the other three models which based on statistical features in BFCN.Because of the following reasons: we utilize multi-granularity to learn the core topology differences and brain salient patterns between MDD and NC in the brain network; The MGGCN model has flexible local receptive fields.Thus, MGGCN model has advantages in MDD recognition effect and model generalization.

C. Ablation Study
To understand the role of multi-granularity in learning the core topology of the brain network, we adopt the three public datasets of MODMA, MPHCE and EDR to measure the MDD recognition effect of single BFCN and multi-granularity The results of ablation study are shown in Table IV, Table V and Table VI.IV, we compare the MDD recognition effect of a single BFCN based on PLV and PLI with that of a multi-granularity.Firstly, we fix the construction method, and detect that multi-granularity is superior to a single BFCN in most metrics of Delta, Theta, Alpha, Beta, Gamma frequency bands and Aver effect.However, the multi-granularity is weaker than that of a single BFCN on the Prec and F1_Score of Theta band in PLV.The reason may be that MGGCN is more sensitive to false negative rate.In terms of Prec, recall and F1_Score metrics on the Delta band, the multi-granularity is weaker than that of a single BFCN in PLI.The reason may be that redundant connection in BFCN leads to better recognition effect.In general, MGGCN utilizes multi-granularity filtering redundant connections to learn the core topology and brain salient patterns, which is better than a single BFCN.
Secondly, fixing a single BFCN or multi-granularity, we detect that a single BFCN based on PLV has stronger Prec, Recall, and F1_Score than PLI on the Beta band in single BFCN; The four metrics of PLV in the Delta band are stronger than those of PLI in multi-granularity.However, PLI are stronger than PLI in its other band and Aver effect.The reason may be that different construction methods have different PS capture sensitivity in different frequency bands.In general, the MDD recognition effect of PLI is better than that of PLV.
2) Result on MPHCE Dataset: As shown in Table V, firstly, we fix the construction method and detect that the four metrics of a single BFCN in the Delta band and the Acc in the Theta band are stronger than multi-granularity in PLV; Prec, Recall, F1_Score of a single BFCN in the Delta band and Prec, Recall, F1_Score of the Beta band are stronger than multi-granularity in PLI.However, the multiple metrics of other frequency bands and Aver effect of multi-granularity are better.Secondly, fixing a single BFCN or multi-granularity, we can detect that PLV based construction methods are better than PLI in Theta and Alpha frequency bands in single BFCN.However, all metrics of all bands based on PLI are better than PLV in multigranularity.In general, the MDD recognition effect of PLI based construction methods is better than that of PLV.
3) Result on EDR Dataset: As shown in Table VI, firstly, we fix the construction method and find that the Acc of a single BFCN in the Gamma band is stronger than that of multi-granularityn in PLV; The Acc of a single BFCN in the Delta and Theta bands is stronger than that of multi-granularity in PLI.However, multi-granularity are better than that of a single BFCN in multiple metrics of the other bands and Aver effect.
Secondly, fixing a single BFCN or multiple granularity, we can detect that PLV based construction methods are better than PLI in Delta frequency band.However, all metrics of all bands based on PLI are better than PLV in multi-granularity.In general, the MDD recognition effect of PLI based construction methods is better than that of PLV.
Based on the ablation study, we can draw the following conclusions, as shown in Tables IV, V and VI.First of all, through the measurement of multiple metrics of different datasets, we can conclude that the MDD recognition effect of MGGCN based on multi-granularity is better than that of a single BFCN.By comparison, we can see that multi-granularity is eliminating the interference of redundant connections on Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.feature extraction.The recognition effect of this model in multiple frequency bands and Aver effect is better than that based on a single BFCN.Secondly, fixing the single model, for example, fixing the single BFCN or multi-granularity, we can draw a conclusion that the MDD recognition effect of PLI is better than PLV, which indicates that PLI is more sensitive than PLV in capturing EEG phase changes.

D. Sensitivity Analysis
1) The Changes of Multi-Granularity: To explore the differences in brain core topology between MDD and NC, we compared and analyzed the core topology of brain networks at different granularity.Due to the different PS intensities captured by different construction methods, the same initial and update granularity is used to ensure the consistency of variables and prevent model MDD based on different construction methods from producing different recognition effects due to different granularity.
MGGCN recognizes MDD by learning the differential core topological structure of brain networks between MDD and NC.To prevent excessive filtering at multiple granularities that might exclude redundant connections with functional relevance or introduce redundant connections due to insufficient granularity, we set the initial granularity to 0.01 based on prior work experience.We then update the granularity value to 0.003 for 10 iterations (as a zero-value network might emerge with further refinement).
As shown in Fig. 4  sub BBNs observed from PLI in the Beta frequency bands of Subject 1 (MDD) and Subject 1 (NC) from the MPHCE dataset.
By comparing the core topological structure of the BDCN between MDD and NC at different granularity, we found that as the granularity increased, the core topological structure of the brain network became more obvious.The connection between the RF and RT regions continues to decrease, while the brain connection between LT and LP continues to decrease, resulting in increasingly obvious brain network connectivity defects.However, the core extension structure of NC's brain network does not have connectivity defects.
From this, we can conclude that the brain functional connections between RF and RT, LP and LP can serve as potential biomarkers for MDD recognition. 1

IV. CONCLUSION AND FUTURE WORK
This paper proposes a multi-granularity graph convolution model (MGGCN) for MDD recognition.Particularly, MGGCN constructs the multi-granularity BFCN in the graph for the first time.MGGCN recognizes MDD and NC by building a multi-granularity neural network to learn the core topology of the brain network.With the advance of the multi-granularity neural network, MGGCN can retain valuable weak connections while filtering redundant connections in BFCN.Experimental results on three benchmark datasets show that MGGCN has a flexible local receptive field and strong generalization ability.In addition, we also conduct an ablation study, and the results show that the multi-granularity method has advantages over the single BFCN method.Our proposed model can accurately and quickly learn the differences in core brain network topology between MDD and NC, and thus recognize MDD.Increasing granularity shows a corresponding 1 BFCN map is drawn with the help of BrainNet Viewer [49].
increase in connectivity defects in the RF and RT regions, as well as the LT and LP regions.These brain functional connections could potentially serve as biomarkers for recognizing MDD.
In future research work, due to the current lack of attention to the differences in subjects with mild, moderate, and severe depression, we tend to pay more attention to the mechanism of depression and its evolution from mild to severe.We will combine AI technology with clinical diagnosis, model the development trend of depression, and study how to control and early intervene in the development of depression.

Fig. 2 .
Fig. 2. Structure of multi-granularity graph convolution network based on brain network core topology.

1 )
Result on MODMA Dataset: As shown in Table

Fig. 4 .
Fig. 4. Comparison of core topological structures of brain networks between MDD and NC based on multi-granularity.
, the scale area of the brain is divided into two hemispheres (left and right); The frontal region (dark blue) is divided into left frontal region (LF), right frontal region (RF); The temporal region (orange) is divided into left temporal region (LT), and right temporal region (RT); The central region(light green) is divided into left central region (LC) and right central region (RC); The posterior region (light blue) is divided into left posterior region (LP) and right posterior region(RP) [48].Fig.4 shows the pyramid of

TABLE VI COMPARISON
OF MDD RECOGNITION EFFECTS (%) OF EDR DATASET