Depression EEG Classification Method Based on Dual-Branch Fusion Model

Abstract

The present invention discloses a method for classifying depressive electroencephalogram based on a dual-branch fusion model in deep learning, comprising the following steps: (1) obtaining electroencephalogram signals of the Fp1, Fpz, and Fp2 electrodes in the prefrontal lobe of the brains of a number of healthy individuals, (2) obtaining electroencephalogram signals of the Fp1, Fpz, and Fp2 electrodes in the prefrontal lobe of the brains of a number of patients with mild depression, (3) obtaining electroencephalogram signals of the Fp1, Fpz, and Fp2 electrodes in the prefrontal lobe of the brains of a number of patients with moderate depression, (4) training and learning the dual-branch fusion model with the input forms of healthy controls, patients with mild depression, and patients with moderate depression in steps (1), (2), and (3), and (5) converting the electroencephalogram signals of the window to be analyzed into corresponding wavelet time-frequency diagrams and inputting them into the dual-branch fusion model trained in step (4) to complete the analysis of the electroencephalogram signals. This method has good effects and can distinguish between depression and health as well as the degree of depression.

Description

Technical Field

[0001] The present invention relates to a method for classifying depressive electroencephalogram based on a dual-branch fusion model, belonging to the field of computer application technology. Background Art

[0002] Major Depression Disorder (MDD) is a common mental disorder. Its clinical symptoms include low interest in all things, weak self-identity, inattention, etc., and even repeated self-harm and suicidal behaviors. According to the estimation of the World Health Organization (WHO), by 2030, the number of patients with depression will exceed the total number of patients with all cardiovascular diseases, and depression will become the leading cause of self-harm in the world. In China, the incidence of depression accounts for about 4.2% of the total population, showing an increasing and younger trend year by year. Depression not only causes serious harm to individuals, but also has a negative impact on the families and society of patients. If patients can be correctly diagnosed at the early stage of depression, their conditions can be significantly improved through means such as psychotherapy, drug treatment, electroconvulsive therapy, and lifestyle changes. Therefore, accurate diagnosis of depression is of great significance.

[0003] On the one hand, electroencephalogram (EEG) signals can capture the millisecond-level neuronal electrical activities of the brain, with high temporal resolution and reliability, and relatively low cost. On the other hand, existing research has shown that the degree of depression of patients is mainly related to the activity of prefrontal EEG and the symmetry of EEG between the left and right frontal lobes. Therefore, three-channel EEG data of the prefrontal lobe of the brain can be selected as the basis for diagnosing depression. Summary of the Invention

[0004] In order to accurately distinguish between healthy and depressive states and the degree of depression, the present invention provides a method for classifying depressive EEG based on dual-branch fusion. This method uses the wavelet time-frequency diagram of three-channel EEG data of the prefrontal lobe and the original EEG sequence as inputs to train a dual-branch fusion model to distinguish between healthy and moderately depressive EEG, as well as between mildly depressive and moderately depressive EEG.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] The present invention provides a method for classifying depressive EEG based on deep learning, including the following steps:

[0007] (1) Obtain the electroencephalogram (EEG) signals of the prefrontal lobe of the brains of several groups of healthy individuals from the Fp1, Fpz, and Fp2 electrodes, and use a sliding window to segment them into window data. Then, convert them into wavelet time-frequency diagrams using wavelet transform. Finally, use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input to the model.

[0008] (2) Obtain the electroencephalogram (EEG) signals of the prefrontal lobe of the brains of several groups of patients with mild depression from the Fp1, Fpz, and Fp2 electrodes, and use a sliding window to segment them into window data. Then, convert them into wavelet time-frequency diagrams using wavelet transform. Finally, use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input to the model.

[0009] (3) Obtain the electroencephalogram (EEG) signals of the prefrontal lobe of the brains of several groups of patients with moderate depression from the Fp1, Fpz, and Fp2 electrodes, and use a sliding window to segment them into window data. Then, convert them into wavelet time-frequency diagrams using wavelet transform. Finally, use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input to the model.

[0010] (4) Use the input forms of the healthy controls, patients with mild depression, and patients with moderate depression in steps (1), (2), and (3) to train and learn the dual-branch fusion model.

[0011] (5) Convert the EEG signals of the window to be analyzed into corresponding wavelet time-frequency diagrams, and input them into the dual-branch fusion model trained in step (4) to complete the analysis of the EEG signals.

[0012] As a further technical solution of the present invention, a universal three-channel EEG acquisition system is used for prefrontal lobe acquisition.

[0013] As a further technical solution of the present invention, the sliding window lengths in (1), (2), and (3) are 2 seconds, and the overlap rate is 0.

[0014] As a further technical solution of the present invention, the wavelet transform used is a wavelet transform with a complex Morlet wavelet (broadband parameter is 3, center wavelength is 3) as the base wavelet.

[0015] As a further technical solution of the present invention, the dual-branch fusion model is a deep learning classification model built based on a convolutional neural network and a channel attention mechanism.

[0016] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solutions and has the following technical effects: The dual-branch fusion model proposed by the present invention can effectively extract the time-frequency features of EEG signals, and uses a convolutional neural network to extract effective information of EEG from two modalities. The accuracy, precision, recall rate, and F1 score in distinguishing healthy and moderately depressed, healthy and mildly depressed, mildly depressed and moderately depressed, and healthy and mildly and moderately depressed are all higher than those of existing EEG methods for depression. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the method disclosed by the present invention;

[0018] Figure 2 It is a dual-branch fusion model designed by the present invention;

[0019] Figure 3 It is a schematic diagram of the positions of the three-channel EEG of the prefrontal lobe selected in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The present invention will be further clarified below in conjunction with the drawings and specific embodiments.

[0021] Embodiment: As Figure 1 shown, it is a flowchart of the method for classifying EEG of depression based on deep learning disclosed by the present invention, which specifically includes the following steps:

[0022] (1) Obtain the EEG signals of the Fp1, Fpz, and Fp2 electrodes of the prefrontal lobe of the brains of a number of healthy people, and use a sliding window to cut them into window data, then use wavelet transform to convert them into wavelet time-frequency diagrams, and finally use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input of the model.

[0023] (2) Obtain the EEG signals of the Fp1, Fpz, and Fp2 electrodes of the prefrontal lobe of the brains of a number of mildly depressed patients, and use a sliding window to cut them into window data, then use wavelet transform to convert them into wavelet time-frequency diagrams, and finally use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input of the model.

[0024] (3) Obtain the EEG signals of the Fp1, Fpz, and Fp2 electrodes of the prefrontal lobe of the brains of a number of moderately depressed patients, and use a sliding window to cut them into window data, then use wavelet transform to convert them into wavelet time-frequency diagrams, and finally use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input of the model.

[0025] Based on the relationship between the prefrontal lobe of the brain and depression, the electroencephalogram (EEG) of Fp1, Fpz, and Fp2 in the prefrontal lobe was selected as the data source, and a universal three-channel EEG acquisition system was used as the acquisition tool. The specific acquisition steps are as follows: I. Measure the distance from the nasion (the midpoint between the eyebrows) to the inion (the prominent bone at the back of the head), and record it as one-tenth of the distance from the midpoint between the eyebrows to the top of the head, denoted as, which is the position of Fpz; II. Measure the distance between the bilateral mastoids over the top of the head, and take one-tenth of the distance from the left mastoid to the top of the head, denoted as point; III. Take one-tenth of the distance from the occipital bone to the top of the head, and take the positions passing through one-tenth of the upper left, upper right, and upper middle respectively, which are the positions of Fp1 and Fp2.

[0026] The collected EEG data is preprocessed according to the following steps to obtain relatively clean EEG data: First, a Butterworth band-pass filter from 0.5 Hz to 100 Hz is used to obtain the effective frequency of the EEG, then the ICA algorithm is used to remove artifacts and a 50 Hz notch filter is used to remove power frequency noise, and finally min-max data normalization is used to prevent the influence of dimensionality and improve the convergence speed of the model.

[0027] Secondly, the following steps are used to convert the EEG data into an input form suitable for the dual-branch fusion model: First, a sliding window is used to intercept the EEG signal: a sliding window with a length of 2 seconds and an overlap rate of 0 is used to intercept the original collected EEG. Then, it is transformed into a wavelet coefficient matrix through wavelet transform with the complex Morlet wavelet (broadband parameter is 3, center wavelength is 3) as the base wavelet. Finally, the modulus of the complex Morlet wavelet coefficient is taken as the pixel value of the image and the size of the image is sampled to 224x224.

[0028] (4) In step (4), the dual-branch fusion model is trained and learned with the input forms of healthy controls, patients with mild depression, and patients with moderate depression in (1), (2), and (3).

[0029] (4-1) The dual-branch fusion model is composed of a picture feature extraction branch and a temporal feature extraction branch.

[0030] (4-2) The temporal feature extraction branch in the dual-branch fusion model embeds an attention mechanism based on discrete Fourier transform coefficients in the FCN model. The FCN branch is composed of three basic blocks (Conv1d+BN+ReLU), where the convolution output channels of the three basic blocks are 128, 256, and 128 respectively. The attention mechanism based on discrete Fourier transform coefficients is an extension of the SE model in the DFT frequency domain.

[0031] (5) Convert the EEG signal of the window to be analyzed into the corresponding wavelet time-frequency diagram, and input it into the dual-branch fusion model trained in step (4) to complete the analysis of the EEG signal.

[0032] Example 2:

[0033] Steps (1), (2), and (3): Collect the electroencephalogram (EEG) of the subjects and divide the degree of depression according to the depression scale, and then preprocess the EEG data. In this embodiment, a universal three-channel EEG acquisition system is used to collect the EEG signals of the prefrontal lobes Fp1, Fpz, and Fp2 of the subjects' brains. The sampling frequency f is 1000 Hz, and each subject is grouped by at least two professional psychiatrists using the Hamilton Depression Rating Scale (HDRS-17). A total of 20 healthy people, 34 patients with mild depression, and 29 patients with moderate depression are divided. Then, a sliding window with a length of 2 s and an overlap rate of 0 is used to segment the three-channel EEG signals of each subject, and they are transformed into a wavelet coefficient matrix through wavelet transform with a complex Morlet wavelet (broadband parameter is 3, central wavelength is 3) as the base wavelet. Then, the modulus of the complex Morlet wavelet coefficients is taken as the pixel value of the image, and the size of the image is sampled to 224×224. Finally, three wavelet time-frequency diagrams and their corresponding original EEG sequences are combined to form an input tensor.

[0034] Step (4), construct a dual-branch fusion model and input the input tensor into the dual-branch fusion model for training. In this embodiment, the parameters of the dual-branch fusion model are as follows in the table. Conv under the image feature extraction branch represents a two-dimensional convolutional layer, while Conv in the feature extraction branch represents a one-dimensional convolutional layer; BN represents batch normalization, ReLU represents the ReLU layer, and MaxPool represents the max pooling layer.

[0035] 1) Feature extraction module 1: Consists of a one-dimensional convolutional layer, a batch normalization layer, a ReLU layer, a Dropout layer, and the dropout rate of the DFTC-ATT layer is 0.3;

[0036] 2) Feature extraction module 2: Structurally, it is exactly the same as feature extraction module 1; in terms of parameters, the size of the convolutional kernel becomes 5, and the number of input and output channels of the convolutional layer becomes 128 and 256.

[0037] 3) Feature extraction module 3: Structurally, compared with feature extraction module 1, GAP is used to replace DFTC-ATT; in terms of parameters, the size of the convolutional kernel becomes 3, and the number of input and output channels of the convolutional layer becomes 256 and 128.

[0038] Finally, the feature vectors learned by the two branches are concatenated into a final feature vector (with a length of 192), and the classification result is input through the classification module.

[0039] Step (5), evaluate the learning effect of the model using accuracy, precision, recall, and F1 score.

[0040] In the present invention, Non-De is denoted as healthy subjects, Mid-De as patients with mild depression, Con-De as patients with moderate depression, TN (True Negative) as negative-class samples predicted as negative by the model, FP (False Positive) as negative-class samples predicted as positive by the model, and FN (False Negative) as positive-class samples predicted as negative by the model. Then, the accuracy is defined as the probability of correct classification of all samples:

[0041]

[0042] Precision can be divided into precision of positive-class samples and precision of negative-class samples. The precision of positive-class samples is the proportion of samples that are actually positive among the samples predicted as positive:

[0043]

[0044] The precision of negative-class samples is the proportion of samples that are actually negative among the samples predicted as negative:

[0045]

[0046] Recall can also be divided into recall of positive-class samples and recall of negative-class samples. The recall of positive-class samples is the proportion of samples that are determined as positive among the samples that are actually positive:

[0047]

[0048] The recall of negative-class samples is the proportion of samples that are determined as negative among the samples that are actually negative:

[0049]

[0050] The F1 value comprehensively considers precision and recall and is the harmonic mean of precision and recall. It is often used as the final evaluation method for machine learning classification methods. The higher the F1 value for each class, the better the classification result. The F1 value for each category is expressed as:

[0051]

[0052] In the present invention, different classification models (ResNet, FCN, MLSTM-FCN, CNN-LSTM) are used for comparison, as shown in Tables 1 to 4.

[0053] Table 1 Comparison of results of different classification methods for Non-De and Con-De tasks

[0054]

[0055] Table 2 Comparison of Results of Different Classification Methods under Non-De and Mid-De Tasks

[0056]

[0057] Table 3 Comparison of Results of Different Classification Methods under Mid-De and Con-De Tasks

[0058]

[0059] Table 5 Comparison of Results of Different Classification Methods under Non-De, Mid-De and Con-De Tasks

[0060]

[0061] From the experimental results in Tables 1 to 5, it can be concluded that the classification effect of the present invention is significantly better than that of other models.

[0062] As described above, it is only the specific implementation manner in the present invention, but the protection scope of the present invention is not limited thereto. Any transformation or replacement that can be understood and conceived by those familiar with the technology within the technical scope disclosed by the present invention should be covered within the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for classifying depressive electroencephalogram based on deep learning of a dual-branch fusion model, characterized in that, The method includes the following steps: (1) Obtain the electroencephalogram signals of the Fp1, Fpz, and Fp2 electrodes in the prefrontal lobe of the brains of a number of healthy individuals, and use a sliding window to segment them into window data. Then, convert them into wavelet time-frequency diagrams using wavelet transform. Finally, use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input to the model; (2) Obtain the electroencephalogram signals of the Fp1, Fpz, and Fp2 electrodes in the prefrontal lobe of the brains of a number of patients with mild depression, and use a sliding window to segment them into window data. Then, convert them into wavelet time-frequency diagrams using wavelet transform. Finally, use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input to the model; (3) Obtain the electroencephalogram signals of the Fp1, Fpz, and Fp2 electrodes in the prefrontal lobe of the brains of a number of patients with moderate depression, and use a sliding window to segment them into window data. Then, convert them into wavelet time-frequency diagrams using wavelet transform. Finally, use the wavelet time-frequency diagrams of the three channels and the corresponding original window sequences as the input to the model; (4) Train and learn the dual-branch fusion model with the input forms of healthy controls, patients with mild depression, and patients with moderate depression in steps (1), (2), and (3); (5) Convert the electroencephalogram signal of the window to be analyzed into the corresponding wavelet time-frequency diagram, and input it into the dual-branch fusion model trained in step (4) to complete the analysis of the electroencephalogram signal; The sliding window is set to have a length of 2 seconds and an overlap rate of 0; Select the wavelet transform algorithm as the wavelet transform with Morlet wavelet as the base wavelet. The broadband parameter of the Morlet wavelet is 3 and the center wavelength is 3; The dual-branch fusion model is a deep learning classification model built based on a convolutional neural network and a channel attention mechanism; Step (4) is specifically as follows: (4-1) The dual-branch fusion model is composed of a picture feature extraction branch and a temporal feature extraction branch; (4-2) The temporal feature extraction branch in the dual-branch fusion model embeds an attention mechanism based on discrete Fourier transform coefficients in the FCN model. The FCN branch is composed of three basic blocks, and the three basic blocks are composed of Conv1d + BN + ReLU. The convolution output channels of the three basic blocks are 128, 256, and 128 respectively. The attention mechanism based on discrete Fourier transform coefficients is an extension of the SE model in the DFT frequency domain.

Images