Method for identifying schizophrenia by utilizing EEGNet model based on electroencephalogram signals

Abstract

The invention belongs to the technical field of electroencephalogram signal processing, and discloses a method for identifying schizophrenia by utilizing an EEGNet model based on electroencephalogram signals, which comprises the following steps of: acquiring the electroencephalogram signals of a subject, preprocessing the electroencephalogram signals, converting time domain signals into frequency domain signals by using Fourier transform on the preprocessed electroencephalogram signals, and then extracting gamma frequency bands; according to a one-leaving cross validation method, dividing the extracted gamma frequency band into a training set and a test set, and dividing a validation set based on the training set; constructing an EEGNet model, and training the EEGNet model through the training set, the test set and the verification set to obtain a trained EEGNet model; and inputting a to-be-detected electroencephalogram signal into the trained EEGNet model to obtain a classification result. The method is suitable for diagnosing whether the patient suffers from schizophrenia or not according to the electroencephalogram signals, the deviation of subjective judgment can be reduced, and the misdiagnosis rate is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram (EEG) signal processing, and relates to a method for identifying schizophrenia by using an EEGNet model based on electroencephalogram (EEG) signals. Background Art

[0002] Schizophrenia (SCZ) is a chronic, serious brain disease that includes auditory hallucinations, delusions, and thinking and behavioral disorders, which impair patients' daily functions and social interactions. In severe cases, patients may also engage in self-harm or suicidal behavior, seriously damaging their mental and physical health and placing a heavy burden on their families and society.

[0003] At present, the screening of SCZ is mainly observed and recorded through questionnaires, psychological scales and some behavioral tests, and the corresponding psychologists or clinicians are combined with the characteristics of electroencephalogram (EEG) to make judgments. As a non-invasive neurophysiological technology with high temporal resolution, low cost and easy operation, EEG has been widely used in schizophrenia research and auxiliary diagnosis. A large number of preclinical studies have shown that in the resting state of SCZ patients, the activity of the gamma frequency band of the brain is significantly increased.

[0004] EEGNet is a deep learning model for processing EEG signals, which has performed well in the field of brain-computer interface. Its compact structure design not only maintains efficiency and accuracy, but also significantly reduces computational costs and model complexity, and has been rapidly developed in the field. Therefore, it is of great significance to provide a method for identifying SCZ data based on the EEGNet architecture through the gamma frequency band features in the resting state EEG signal. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying schizophrenia based on EEG signals using the EEGNet model, which can quickly obtain diagnostic information of the EEG signals to be tested, thereby reducing the bias of subjective judgment, reducing the misdiagnosis rate, and providing a scientific basis for the formulation of personalized treatment plans.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying schizophrenia based on EEG signals using an EEGNet model comprises the following steps:

[0008] S1, collecting the EEG signal of the subject, and preprocessing the EEG signal to obtain the preprocessed EEG signal;

[0009] S2, using Fourier transform on the preprocessed EEG signal, converting the time domain EEG signal into a frequency domain EEG signal, and then extracting the γ frequency band;

[0010] S3, according to the leave-one-out cross-validation method, the extracted γ frequency band is divided into a training set and a test set, and the validation set is divided based on the training set;

[0011] S4, build an EEGNet model, train the EEGNet model through the training set, find the optimal EEGNet model after training through the validation set, and finally evaluate the performance of the optimal EEGNet model after training through the test set to obtain the trained EEGNet model;

[0012] S5. Input the EEG signal to be tested into the trained EEGNet model to obtain the classification result.

[0013] As a limitation, in step S1, the process of preprocessing the EEG signal is:

[0014] S11, remove irrelevant EEG signal collection electrodes, including horizontal eye movement electrode HEOG, vertical eye movement electrode VEOG, CB1 electrode and CB2 electrode;

[0015] S12, re-reference, using bilateral mastoid electrodes M1 and M2 as reference electrodes;

[0016] S13, apply notch filter, eliminate 50Hz power interference through EEGLAB, and use 0.5~80Hz bandpass filter;

[0017] S14, data segmentation, segmenting each subject’s EEG signal into 2-second time segments and discarding time segments containing significant artifacts;

[0018] S15, independent component analysis, removes artifact-related components, and then downsamples the EEG signal to 500 Hz.

[0019] As a second limitation, in step S4, the EEGNet model includes a convolution module, a depth convolution module, a separable convolution module and a classification module connected in sequence;

[0020] The convolution module includes a Conv2D layer and a first BatchNormalization layer connected in sequence;

[0021] The deep convolution module includes a DepthwiseConv2D layer, a second BatchNormalization layer, a first ELU layer, a first AveragePooling2D layer, and a first Dropout layer connected in sequence;

[0022] The separable convolution module includes a SeparableConv2D layer, a third BatchNormalization layer, a second ELU layer, a second AveragePooling2D layer, and a second Dropout layer connected in sequence;

[0023] The classification module includes a Flatten layer, a Dense layer, and a softmax layer connected in sequence.

[0024] As a limitation, eight two-dimensional convolutional filters of size (1, 64) are used in the Conv2D layer of the convolution module;

[0025] The size of the DepthwiseConv2D layer of the deep convolution module is (60, 1) and the depth is 2; the first AveragePooling2D layer of the deep convolution module is the average pooling of (1, 4); the dropout rate of the first Dropout layer of the deep convolution module is 0.5;

[0026] The size of the SeparableConv2D layer of the separable convolution module is (1, 16); the second AveragePooling2D layer of the separable convolution module is average pooling of (1, 8).

[0027] Due to the adoption of the above technical solution, the present invention has the following technical advances compared with the prior art:

[0028] (1) The present invention improves the accuracy, timeliness and objectivity of schizophrenia diagnosis by extracting the gamma frequency band of the EEG signal and then using the gamma frequency band to train the EEGNet model;

[0029] (2) The present invention classifies the EEG signals to be tested through the trained EEGNet model, which reduces the workload of professional neurologists in data interpretation, and also reduces the deviation introduced by human factors, thereby improving the accuracy and reliability of diagnosis;

[0030] (3) The present invention has development prospects in the field of schizophrenia identification, has high accuracy, and is easy to promote and apply.

[0031] In summary, the present invention is suitable for diagnosing whether a patient suffers from schizophrenia based on EEG signals, which can reduce the deviation of subjective judgment and lower the misdiagnosis rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Shown is a method flow chart of an embodiment of the present invention;

[0033] Figure 2 Schematic diagram of leave-one-out cross-validation method according to an embodiment of the present invention;

[0034] Figure 3 Shown is a training architecture diagram of the EEGNet model according to an embodiment of the present invention;

[0035] Figure 4 The figure shows the average results of the training and validation sets based on 28-fold cross validation in the embodiment of the present invention;

[0036] Figure 5 Shown is a comparison of classification accuracy between schizophrenia subjects and healthy control subjects in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0038] Embodiment A method for identifying schizophrenia based on EEG signals using EEGNet model

[0039] like Figure 1 As shown, this embodiment is a method for identifying schizophrenia based on EEG signals using an EEGNet model, comprising the following steps:

[0040] S1, collecting the EEG signal of the subject, and preprocessing the EEG signal to obtain the preprocessed EEG signal;

[0041] Among them, the process of preprocessing the EEG signal is:

[0042] S11, remove irrelevant EEG signal collection electrodes, including horizontal eye movement electrode HEOG, vertical eye movement electrode VEOG, CB1 electrode and CB2 electrode;

[0043] S12, re-reference, using bilateral mastoid electrodes M1 and M2 as reference electrodes;

[0044] S13, apply notch filter, eliminate 50Hz power interference through EEGLAB, and use 0.5~80Hz bandpass filter;

[0045] S14, data segmentation, segmenting each subject’s EEG signal into 2-second time segments and discarding time segments containing significant artifacts;

[0046] S15, independent component analysis, to remove artifact-related components, and then downsample the EEG signal to 500 Hz;

[0047] S2, using Fourier transform on the preprocessed EEG signals, converting the time domain EEG signals into frequency domain EEG signals, and then extracting the γ frequency band (30-80Hz);

[0048] In this step, since the γ frequency band is closely related to higher cognitive functions, it has been found to be associated with the pathological mechanism of schizophrenia; therefore, this step can improve the accuracy, timeliness and objectivity of diagnosis by extracting and analyzing the γ frequency band;

[0049] S3, according to the leave-one-out cross-validation method, the extracted γ frequency band is divided into a training set and a test set, and the validation set is divided based on the training set;

[0050] In this step, the training set and the test set are divided according to the leave-one-out cross-validation method (LOSOCV) to comprehensively evaluate the generalization ability of the model and maximize the use of data. In this method, the EEG signal data set of each subject is used as a test set independently, while the EEG signal data of all other subjects are used for training; in addition, 30% of the training set is divided into a validation set, through which the performance of the model can be monitored during the training process, and the learning rate can be dynamically adjusted according to the loss rate or accuracy change on the validation set; the current optimal model can also be saved through the validation set. After the training is completed, the model with the best performance in the validation set is loaded for the final evaluation of the test set; the loss rate and accuracy recorded by the validation set can be compared with the indicators of the training set to analyze whether the model is underfitting or overfitting;

[0051] like Figure 2 The figure shows a schematic diagram of the leave-one-out cross-validation method of this embodiment. The EEG signal data of the subject is divided into 28 folds, and each fold corresponds to a complete EEG signal data set of a subject. This method is suitable for small data sets because it maximizes the size of the training set, and at the same time, the EEG signal data of each subject is evaluated separately, which effectively evaluates the adaptability of the model to individual differences. Therefore, the leave-one-out cross-validation method is an effective evaluation scheme that can demonstrate the robustness of the classifier. In order to optimize the model, this step uses the Adam optimizer to fit the model and optimizes the performance by minimizing the category cross entropy loss function. In addition, the ReduceLROnPlateau scheduler is introduced, which can automatically adjust the learning rate according to the change of the validation set loss, thereby improving the convergence of the model. After completing 250 training iterations (epochs), the model weights that perform best on the validation set are saved to ensure the efficient performance of the model.

[0052] S4, build an EEGNet model, train the EEGNet model through the training set, find the optimal EEGNet model after training through the validation set, and finally evaluate the performance of the optimal EEGNet model after training through the test set to obtain the trained EEGNet model;

[0053] like Figure 3The training architecture diagram of the EEGNet model in this step is shown in Table 1. The parameters of the EEGNet model architecture are shown in Table 1. As can be seen from Table 1, the EEGNet model first formats the input through an initial shaping layer. The EEGNet model includes a convolution module, a depth convolution module, a separable convolution module and a classification module connected in sequence;

[0054] Table 1

[0055]

[0056] The convolution module includes a Conv2D layer and a first BatchNormalization layer connected in sequence. In the Conv2D layer of the convolution module, eight two-dimensional convolution filters of size (1, 64) are used to generate eight feature maps for capturing brain activity patterns across different frequency ranges. In order to stabilize the training process and reduce overfitting, the first BatchNormalization layer is added to the convolution module.

[0057] The deep convolution module includes a DepthwiseConv2D layer, a second BatchNormalization layer, a first ELU layer, a first AveragePooling2D layer and a first Dropout layer connected in sequence; in order to avoid the DepthwiseConv2D layer being fully connected to all previous feature maps, this step sets the size of the DepthwiseConv2D layer of the deep convolution module to (60, 1) and the depth to 2. The DepthwiseConv2D layer of the deep convolution module can learn spatial filters and significantly reduce the number of trainable parameters; then the feature map dimension is batch normalized through the second BatchNormalization layer, and the exponential linear unit ELU is used as a nonlinear activation function in the first ELU layer to enhance the nonlinear processing capability of the network, and the first AveragePooling2D layer with a size of (1, 4) is used to reduce the feature map dimension. Since the data set is used for internal classification of subjects, the deep convolution module adds a first Dropout layer with a drop rate of 0.5 to prevent over-fitting;

[0058] The separable convolution module includes a SeparableConv2D layer, a third BatchNormalization layer, a second ELU layer, a second AveragePooling2D layer, and a second Dropout layer connected in sequence; in the separable convolution module, the size of the SeparableConv2D layer is (1, 16), followed by point-by-point convolution. This design significantly reduces the number of parameters that need to be fitted, and at the same time effectively decouples the internal and mutual relationships of the feature maps by first performing local operations on each feature map and then globally integrating point by point; since different feature maps can represent useful information in different frequency ranges, this architecture is very suitable for processing EEG frequency domain signal data. In addition, the second AveragePooling2D layer of the separable convolution module is an average pooling of (1, 8), which can be used to further reduce the data dimension;

[0059] The classification module includes a Flatten layer, a Dense layer, and a softmax layer connected in sequence; in the classification module, the multidimensional features are simplified into a one-dimensional form and then input into a softmax classifier with two units to distinguish between healthy and schizophrenia categories;

[0060] like Figure 4 The figure shows the average results of the loss and accuracy of the training and validation sets based on 28-fold cross validation. It can be seen from the figure that in the early stage of EEGNet model training, the training loss rate and validation loss rate of the EEGNet model both dropped rapidly, and the training accuracy and validation accuracy increased rapidly. After 50 training iterations, the rate of decline of the training loss rate and validation loss rate gradually slowed down, and stabilized after 200 iterations, and the performance on the validation set was close to 100% accuracy;

[0061] S5. Input the EEG signal to be tested into the trained EEGNet model to obtain the classification result, accurately identify whether it is a normal person or a schizophrenia patient according to the classification result, and display it on the graphical user interface (GUI) interactive system.

[0062] like Figure 5 The figure shows the comparison of classification accuracy of schizophrenia patients and healthy control group HC subjects in this embodiment, wherein the first 14 subjects are schizophrenia patients, and subjects 15-28 are healthy control group HC. The results show that this embodiment has a very high recognition accuracy when identifying schizophrenia patients, and can almost confirm 100% of the patients, but when identifying the healthy control group HC, most of the data of two healthy groups are identified as schizophrenia data, but overall, it still maintains a high recognition accuracy.

[0063] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the above embodiments, those skilled in the art can still modify the technical solutions described in the above embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying schizophrenia based on EEG signals using an EEGNet model, characterized in that: The following steps are involved: S1, collecting the EEG signal of the subject, and preprocessing the EEG signal to obtain the preprocessed EEG signal; S2, using Fourier transform on the preprocessed EEG signal, converting the time domain EEG signal into a frequency domain EEG signal, and then extracting the γ frequency band; S3, according to the leave-one-out cross-validation method, the extracted γ frequency band is divided into a training set and a test set, and the validation set is divided based on the training set; S4, build an EEGNet model, train the EEGNet model through the training set, find the optimal EEGNet model after training through the validation set, and finally evaluate the performance of the optimal EEGNet model after training through the test set to obtain the trained EEGNet model; S5. Input the EEG signal to be tested into the trained EEGNet model to obtain the classification result.

2. The method for identifying schizophrenia based on EEG signals using the EEGNet model according to claim 1, characterized in that: In step S1, the process of preprocessing the EEG signal is as follows: S11, remove irrelevant EEG signal collection electrodes, including horizontal eye movement electrode HEOG, vertical eye movement electrode VEOG, CB1 electrode and CB2 electrode; S12, re-reference, using bilateral mastoid electrodes M1 and M2 as reference electrodes; S13, apply notch filter, eliminate 50Hz power interference through EEGLAB, and use 0.5~80Hz bandpass filter; S14, data segmentation, segmenting each subject’s EEG signal into 2-second time segments and discarding time segments containing significant artifacts; S15, independent component analysis, removes artifact-related components, and then downsamples the EEG signal to 500 Hz.

3. The method for identifying schizophrenia based on EEG signals using the EEGNet model according to claim 1, characterized in that: In step S4, the EEGNet model includes a convolution module, a depth convolution module, a separable convolution module and a classification module connected in sequence; The convolution module includes a Conv2D layer and a first BatchNormalization layer connected in sequence; The deep convolution module includes a DepthwiseConv2D layer, a second BatchNormalization layer, a first ELU layer, a first AveragePooling2D layer, and a first Dropout layer connected in sequence; The separable convolution module includes a SeparableConv2D layer, a third BatchNormalization layer, a second ELU layer, a second AveragePooling2D layer, and a second Dropout layer connected in sequence; The classification module includes a Flatten layer, a Dense layer, and a softmax layer connected in sequence.

4. The method for identifying schizophrenia based on EEG signals using the EEGNet model according to claim 3, characterized in that: The Conv2D layer of the convolution module uses eight two-dimensional convolution filters of size (1, 64); The size of the DepthwiseConv2D layer of the deep convolution module is (60, 1) and the depth is 2; the first AveragePooling2D layer of the deep convolution module is the average pooling of (1, 4); the dropout rate of the first Dropout layer of the deep convolution module is 0.5; The size of the SeparableConv2D layer of the separable convolution module is (1, 16); the second AveragePooling2D layer of the separable convolution module is average pooling of (1, 8).

Images