An electroencephalogram signal classification method based on graph-node convolutional neural network

Abstract

The application discloses a kind of electroencephalogram signal classification method based on graph-node convolutional neural network.The application uses 2D convolution+graph convolution, graph classification+node classification deep learning network to analyze electroencephalogram signal, its characteristics are: after the original electroencephalogram signal is pretreated, such as denoising, differential entropy feature of electroencephalogram signal is extracted, frequency domain information of electroencephalogram signal is utilized;Then the feature is sent into parallel 2D convolution and graph convolution network to extract time information and spatial information, to obtain the composite feature of time domain, frequency domain and spatial domain, finally input into node classification graph neural network to classify electroencephalogram signal.According to the above method process, the application carries out relevant experimental verification on MODMA data set, and the result proves that the method proposed in the application is better than the method of single model and serial model, and has better electroencephalogram signal classification effect.

Description

Technical Field

[0001] This invention belongs to the field of medical signal processing, specifically relating to a method for classifying electroencephalogram (EEG) signals based on a graph-node-convolutional neural network (GNCNN). Background Technology

[0002] Depression is an impairment in an individual's emotional regulation, clinically manifested as low mood, slowed behavior, and excessive focus on oneself. It not only causes mental anguish but also gradually isolates the individual from the outside world, posing a significant threat to the patient. In clinical practice, objective and accurate identification of patients with depression and timely treatment are prerequisites for recovery.

[0003] Currently, the most common method for classifying depression is a combination of doctor interviews and questionnaire surveys. However, this method has many problems. First, the design criteria for the questionnaires require scores to reach a certain threshold, which often leads to over- or under-reporting results. Second, subjective concealment and other interfering factors cannot be avoided during the interview process. Furthermore, the severe shortage of mental health professionals, differences in doctor experience, patients' lack of attention, and concerns about privacy also contribute to poor classification of depression. Therefore, improving the accuracy and efficiency of depression classification is of paramount importance.

[0004] Although electroencephalography (EEG) technology emerged in the latter half of the 20th century, decades ago only clinicians and neuroscientists were able to collect and analyze EEG data using electrodes with 64 or more channels under strictly controlled laboratory conditions. These high-density electrodes, evenly distributed on the scalp, were used to help researchers discover potential neural mechanisms related to motor, cognitive, or emotional processing. In recent years, with the development of computer hardware and processor technology, our understanding of the brain has deepened, leading to a more profound understanding of its structure and activity patterns. EEG-based products have begun to be applied in various fields of production and daily life.

[0005] Electroencephalography (EEG) is most commonly used in the medical field using methods based on manual feature extraction combined with various classifiers, such as time-domain features, frequency-domain features, time-frequency-domain features, other nonlinear features, and deep learning-based feature extraction. Compared with traditional EEG signal processing methods, deep learning-based methods have been widely used in EEG signal processing and have achieved good results. In recent years, EEG-based and deep learning-based methods have also achieved good performance in the field of medical signal processing, but the classification accuracy of EEG signals still needs to be improved. Summary of the Invention

[0006] To improve the accuracy of depression identification and better facilitate clinical applications, this invention proposes a brainwave signal classification method based on graph-node convolutional neural networks.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for classifying electroencephalogram (EEG) signals based on a graph-node convolutional neural network is proposed. This method first manually extracts the differential entropy features of the EEG signals. Then, the extracted features are input into a parallel 2D convolutional and graph convolutional network structure for deep feature extraction using a graph classification approach. Finally, the extracted features are input into the graph convolutional network for node classification. The specific steps are as follows:

[0009] (I) Data preprocessing: The raw EEG signal is first bandpass filtered to remove high-frequency noise, eliminate slow drift and power line noise; then, independent component analysis is performed to remove electrooculography interference; then, rereference is performed; finally, the data is segmented and the segment containing the image stimulus is extracted based on the event information in the data as the EEG signal for subsequent processing.

[0010] (II) Feature extraction: Extract differential entropy features from the preprocessed EEG signals in step (I);

[0011] (III) Constructing a deep feature extraction network based on parallel 2D convolution and graph convolution to perform deep feature extraction from a spatiotemporal perspective: First, the differential entropy features of consecutive experiments are superimposed and enhanced. The feature matrices of n experiments are used to obtain m enhanced experimental feature matrices. For the enhanced feature matrices, the feature matrices of every k consecutive enhanced experiments are concatenated to form context features. Then, these features are respectively input into a 2D convolutional network to extract temporal information and into a graph convolutional network to extract spatial information. The data graphing method input into the graph convolutional network is as follows: the EEG channels are used as nodes in the graph, and each experiment corresponds to one graph. The features extracted by the 2D convolutional network and the graph convolutional network are concatenated and fused through a fully connected layer to finally obtain a highly aggregated feature vector.

[0012] (iv) Node classification: m feature vectors of the enhancement experiment are finally extracted from the EEG signal of each subject, which together form a large graph. Each experiment is a node in the graph, and the node features are the feature vectors extracted in step (iii). All nodes of the same subject are fully connected, while nodes of different subjects are not connected. The large graph is fed into a graph convolutional network for node classification to obtain the classification results.

[0013] The beneficial effects of this invention are: designing a mode for parallel feature extraction using 2D convolution and graph convolution to achieve deep feature extraction from a spatiotemporal perspective; achieving high aggregation of features of the same experiment through graph classification; achieving information transfer between different experiments of the same subject through node classification; and improving classification accuracy by aggregating information from neighboring experiments. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of the EEG signal classification method of the present invention;

[0015] Figure 2 This is a flowchart of the preprocessing in this invention;

[0016] Figure 3 This is a network structure diagram of the EEG signal classification method based on the GNCNN model in this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and examples. The overall process of the present invention is as follows: Figure 1 As shown.

[0018] First, this invention preprocesses the raw electroencephalogram (EEG) signals. For example... Figure 2 As shown, the specific process is as follows:

[0019] 1) First, the raw EEG signal is filtered from 1-45Hz. Low-pass filtering removes high-frequency noise; high-pass filtering eliminates slow drift; notch filtering removes power line noise.

[0020] 2) Next, independent component analysis is performed on the filtered signal. First, the signal component topography map of the first subject is drawn. Based on the visualization results, the components that are distributed at the front of the scalp topography map and whose component ranking is relatively high are identified as the electrooculogram components that need to be removed. Then, the electrooculogram components in the first subject are used as templates to find the electrooculogram components in the data of other subjects, thereby removing the electrooculogram components in all subjects.

[0021] 3) Then, a rereference step is performed. Common reference electrodes include bipolar reference, REST reference, and average reference. In this invention, the average reference is selected.

[0022] 4) Finally, the data is segmented. Each trial includes image stimulation. Based on the event information in the data, the segment containing the image stimulation is extracted as the EEG signal for subsequent processing. In this embodiment, the data is segmented according to the start and end times of each trial. Each segment contains a complete trial. The data from the 0.5s time period during which the subject is given image stimulation in each trial is extracted as a segment. There are a total of 160 trials for each subject, so a total of 160 data segments are extracted.

[0023] Second, the present invention extracts differential entropy features from the preprocessed EEG signal according to formula (1).

[0024]

[0025] During feature extraction, differential entropy features of different frequency bands are extracted for each signal segment. The frequency bands are divided into: 1-4Hz, 2-6Hz, 4-8Hz, 6-11Hz, 8-14Hz, 11-22Hz, 14-31Hz, 22-40Hz, and 31-45Hz, for a total of 9 frequency bands. Therefore, step (II) extracts 9 differential entropy features for each experiment.

[0026] Third, construct a deep feature extraction network based on parallel 2D convolution and graph convolution. For example... Figure 3 As shown, this invention proposes a feature extraction module combining a 2D convolutional network and a graph convolutional network. Compared to single convolutional neural networks and graph neural networks, it has three main advantages: firstly, the input features utilize differential entropy features to extract information in the frequency domain; secondly, the 2D convolutional layer setup can fully extract contextual information between different experiments; and thirdly, the graph convolutional network setup can fully utilize spatial information. Through training the adjacency matrix, it can flexibly connect highly correlated channels, thereby aggregating neighbor information from different channels to the current node, achieving a spatial information aggregation effect. Therefore, this model achieves full fusion and utilization of the temporal, frequency, and spatial features of the experiments, ultimately extracting representative 8-dimensional features. In this embodiment, the differential entropy features of two consecutive experiments are first superimposed for enhancement, resulting in 80 enhanced feature matrices from 160 experimental feature matrices. Each enhanced feature matrix is ​​then concatenated into three consecutive segments to form contextual features. These features are then input into a 2D convolutional network to extract temporal information and a graph convolutional network to extract spatial information. The graph convolutional network uses a separate graph for each experiment, treating each EEG channel as a node, with one graph per experiment. The features extracted by the 2D and graph convolutional networks are concatenated and fused through a fully connected layer to obtain an 8-dimensional feature. LeakyReLU is used as the activation function after the fully connected layer. A softmax layer is then used for binary classification training to improve the performance of the feature extraction module. After training, the input result of the penultimate fully connected layer is used as the model's feature vector, representing the 128-channel × 9 differential entropy feature matrix for each experiment, resulting in an 8-dimensional feature vector after training.

[0027] Fourth, node classification: Eight features are input into a two-layer graph convolutional neural network for node classification to obtain the classification results of the EEG signals, and ablation experiments are conducted based on the innovations proposed in this invention. Eighty enhancement experiments are ultimately extracted from the EEG signals of each subject. All experiments from all subjects constitute a large graph, with each experiment serving as a node in the graph. The node features are the eight-dimensional features extracted in step (III). All nodes within the same subject are fully connected, while nodes between different subjects are unconnected. The large graph is then fed into a two-layer graph convolutional network for node classification to obtain the classification results of the EEG signals. Specifically, the first layer of the graph convolutional network expands the eight-dimensional feature vector to 32 dimensions, and the second layer reduces the 32-dimensional features to 2 dimensions, which is the classification result.

[0028] The dataset used in this experiment is the MODMA dataset, which contains EEG data from 53 participants, including 24 patients with depression and 29 healthy controls. The experimental task is to perform binary classification of the EEG signals of patients with depression and healthy controls.

[0029] The network was trained using the Adam optimizer with an initial learning rate of 0.001, a training time of 20 epochs, a batch size of 64, and a Dropout parameter of 0.7. Six-fold cross-validation was performed. The first five folds contained EEG data from four patients with depression and five healthy controls, and the sixth fold contained EEG data from four patients with depression and four healthy controls. The average classification accuracy of each fold was used as the final classification accuracy.

[0030] Experimental results show that the classification accuracy of the method of this invention is superior to other network models. The results are compared in Table 1, where GNCNN's classification result of 0.75645 is better than other experimental results.

[0031] Table 1

[0032] EEG classification methods accuracy Serial 2D Convolutional and Graph Convolutional Networks 0.5931 Parallel 2D Convolutional and Graph Convolutional Networks 0.6281 Node classification graph convolutional network 0.64143 Serial 2D Convolutional and Graph Convolutional Networks + Node Classification Graph Convolutional Network 0.71902 GNCNN 0.75645

Claims

1. A method for classifying electroencephalogram (EEG) signals based on a graph-node convolutional neural network. This method first manually extracts the differential entropy features of the EEG signals, then inputs the extracted features into a parallel 2D convolutional and graph convolutional network structure for deep feature extraction using a graph classification approach. Finally, the extracted features are input into the graph convolutional network for node classification. The specific steps are as follows: (I) Data preprocessing: The raw EEG signal is first bandpass filtered to remove high-frequency noise, eliminate slow drift and power line noise; then, independent component analysis is performed to remove electrooculography interference; then, rereference is performed; finally, the data is segmented and the segment containing the image stimulus is extracted based on the event information in the data as the EEG signal for subsequent processing. (II) Feature extraction: Extract differential entropy features from the preprocessed EEG signals in step (I); (III) Constructing a deep feature extraction network based on parallel 2D convolution and graph convolution for deep feature extraction from a spatiotemporal perspective: First, the differential entropy features of consecutive experiments are superimposed and enhanced. The feature matrices of n experiments yield m enhanced experiment feature matrices. For the enhanced feature matrices, the feature matrices of every k consecutive enhanced experiments are concatenated to form context features. These are then input into a 2D convolutional network to extract temporal information and a graph convolutional network to extract spatial information. The data input into the graph convolutional network is constructed as follows: the EEG channels are used as nodes in the graph, and each experiment corresponds to one graph; the features extracted by the 2D convolutional network and the graph convolutional network are concatenated, and feature extraction and fusion are performed through a fully connected layer to finally obtain a highly aggregated feature vector. (iv) Node classification: m feature vectors of the enhancement experiment are finally extracted from the EEG signal of each subject, which together form a large graph. Each experiment is a node in the graph, and the node features are the feature vectors extracted in step (iii). All nodes of the same subject are fully connected, while nodes of different subjects are not connected. The large graph is fed into a graph convolutional network for node classification to obtain the classification results.

Images