Parkinson's disease patient electroencephalogram signal classification method based on deep learning

Abstract

The invention relates to the technical field of electroencephalogram signal processing and deep learning, and discloses a Parkinson's disease patient electroencephalogram signal classification method based on deep learning, and the method comprises the steps: collecting an electroencephalogram signal of a Parkinson's disease patient, and then carrying out the preprocessing to obtain an electroencephalogram signal segment; carrying out noise-assisted multi-dimensional empirical mode decomposition on the electroencephalogram signal segments to obtain intrinsic mode components of different frequency bands, and then extracting a phase lag index feature matrix; besides, a Shannon entropy feature matrix is extracted from the electroencephalogram signal segments, a matrix obtained after fusion of the Shannon entropy feature matrix and a phase lag index feature matrix is used as input of a CA-ResFireNet network trained offline, and a classification result and confidence of the electroencephalogram signals are obtained. According to the method, the noise-assisted multi-dimensional empirical mode decomposition technology is utilized, the limitation of Fourier transform and wavelet transform in processing non-stationary and complex electroencephalogram signals is overcome, rich information of the electroencephalogram signals is reflected more comprehensively through multi-level feature extraction, and the classification precision can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of electroencephalogram signal processing and deep learning, and specifically relates to a method for classifying electroencephalogram signals of Parkinson's disease patients based on deep learning. Background Art

[0002] Parkinson's disease (PD), as a neurodegenerative disease, is characterized by a variety of motor and non-motor symptoms, such as resting tremor, bradykinesia, abnormal posture, memory loss, anxiety, and sleep disorders. These symptoms seriously affect the daily life of patients, leading to frequent problems such as unsteady gait and poor mental state. Given that the pathological mechanism of PD has not been fully elucidated yet and there is a lack of systematic early diagnosis and treatment programs, currently, neurology experts mainly rely on the Unified Parkinson's Disease Rating Scale (UPDRS) to evaluate patients. However, this method is limited by the subjective feedback of patients and has deficiencies in time resolution and accuracy. This makes it difficult for neurologists to accurately capture the short-term changes in the patient's condition and formulate targeted and effective treatment plans, thus affecting the treatment effect.

[0003] Electroencephalogram signals have shown extensive application potential in fields related to the nervous system, covering multiple cognitive functions such as attention, memory, and emotion management, and have been widely used in the detection of nervous system diseases. Research has shown that by analyzing specific features in electroencephalogram signals, it is possible to effectively distinguish between Parkinson's disease patients and healthy individuals. In addition, the non-invasive feature of electroencephalogram signals gives it significant advantages in diagnosis, without the need for surgery or other medical interventions, thus being able to more objectively and quantitatively reflect the disease process and symptoms of Parkinson's disease patients. This characteristic provides a good foundation for improving the accuracy of disease monitoring and the patient experience. Using a deep learning network to intelligently classify electroencephalogram signals and then performing manual recognition can effectively improve the efficiency of electroencephalogram signal image reading and increase the recognition accuracy. For example, in the invention with the patent number CN 117224148 A, "A Method for Identifying Parkinson's Disease from Electroencephalogram Signals Based on Frequency Domain and Spatiotemporal Features", this method first extracts the average power features of the Delta, Theta, and Alpha frequency bands in multi-channel electroencephalogram signals and splices them into a one-dimensional vector, which is then sent to an SVM classifier for classification. Secondly, a compact convolutional neural network based on spatiotemporal features is constructed to iteratively extract the spatiotemporal features of electroencephalogram signals, and at the same time, a channel attention mechanism is introduced to adaptively correct the output feature channels. Finally, the optimal weights of the two models are searched within the interval [0,1], and the two models are fused according to the weights to obtain the final classification result. However, this invention focuses on the time-frequency features of electroencephalogram signals and only extracts the average power features of three frequency bands, obtaining relatively single signal features. Therefore, it is necessary to optimize the method for intelligently classifying electroencephalogram signals through deep learning techniques. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for classifying electroencephalogram (EEG) signals of Parkinson's disease patients based on deep learning, which can achieve high-precision analysis of early EEG signals of Parkinson's disease in the time and frequency domains and perform multi-level feature extraction, so as to realize the classification of EEG signals of Parkinson's disease patients.

[0005] To solve the above technical problem, the present invention provides a method for classifying EEG signals of Parkinson's disease patients based on deep learning, including:

[0006] S1. Collect the EEG signals of Parkinson's disease patients, and then perform preprocessing to obtain EEG signal segments;

[0007] S2. After the EEG signal segments are decomposed by noise-assisted multi-dimensional empirical mode decomposition to obtain intrinsic mode components in different frequency bands, then extract the phase lag index feature matrix;

[0008] S3. Extract the Shannon entropy feature matrix from the EEG signal segments, and then fuse the Shannon entropy feature matrix and the phase lag index feature matrix into a Shannon entropy and phase lag index fusion matrix;

[0009] S4. The Shannon entropy and phase lag index fusion matrix passes through the offline-trained CA-ResFireNet network to obtain the classification result and confidence level of the EEG signal.

[0010] As an improvement to the method for classifying EEG signals of Parkinson's disease patients based on deep learning of the present invention:

[0011] The preprocessing includes band-pass filtering, downsampling, artifact removal, and segmentation processing.

[0012] As a further improvement to the method for classifying EEG signals of Parkinson's disease patients based on deep learning of the present invention:

[0013] The noise-assisted multi-dimensional empirical mode decomposition is specifically as follows:

[0014] (1). Generate Gaussian white noise signals:

[0015]

[0016] where n i (t), i = 1, 2, …, s are white noise signals, and T is the signal length;

[0017] (2). Concatenate the Gaussian white noise signals and the EEG signal segments into a multi-component signal:

[0018]

[0019] where x i(t), where \(i = 1, 2, \ldots, v\) are the electroencephalogram (EEG) signals of \(v\) channels, representing the EEG signal segment;

[0020] (3) Decompose the multivariate signal through the multivariate empirical mode decomposition algorithm to generate \(M\) intrinsic mode functions

[0021]

[0022] where \(c\) m (t) is the intrinsic mode function corresponding to the EEG signal segment, \(a\) m (t) is the intrinsic mode function corresponding to the Gaussian white noise signal, and \(r\) M (t) is the remaining residue;

[0023] (4) Discard the intrinsic mode function \(a\) in formula (3) m (t):

[0024]

[0025] where \(IMF(t)\) represents the intrinsic mode functions of different frequency bands obtained after the noise-assisted multi-dimensional empirical mode decomposition of the EEG signal segment.

[0026] As a further improvement of the EEG signal classification method for Parkinson's disease patients based on deep learning in the present invention:

[0027] The process of extracting the Shannon entropy feature matrix is as follows:

[0028] (1) Calculate the Shannon entropy of the single-channel EEG signal:

[0029]

[0030] where \(p\) represents the probability of the one-dimensional EEG signal, represents the \(j\)-th sample of the EEG signal \(x\) i (t) of the \(i\)-th channel;

[0031] (2) Construct the Shannon entropy feature matrix \(ShEn\) f :

[0032] \(ShEn\) f (i, j) = \(ShEn\) s (i) + \(ShEn\) s ′(j) (6)

[0033] where \(ShEn\) f (i, j) represents the element in the \(i\)-th row and \(j\)-th column of the Shannon entropy feature matrix, \(ShEn\) s (i) represents the Shannon entropy of the \(i\)-th channel, \(ShEn\)s ' is the transpose of ShEn s .

[0034] As a further improvement to the method for classifying electroencephalogram signals of Parkinson's disease patients based on deep learning in the present invention:

[0035] The process of extracting the phase lag index feature matrix is as follows:

[0036] Take the first four of the intrinsic mode components IMF(t) to form a feature matrix of size 4*63*250. For each component matrix of size 63*250, calculate the phase lag index between electroencephalogram channels in sequence. The formula is as follows:

[0037]

[0038] where N represents the time point, and Φ rel represents the phase difference between the signals of two channels at time t n , and sign is the sign function;

[0039] After calculating the phase lag index for the four component matrices, a phase lag index feature matrix of size 4*63*63 is obtained.

[0040] As a further improvement to the method for classifying electroencephalogram signals of Parkinson's disease patients based on deep learning in the present invention:

[0041] The CA-ResFireNet network is an improvement based on the SqueezeNet network, including three groups of feature extraction units, a global pooling layer, and a Softmax layer. The first group of feature extraction units includes downsampling and three Fire_DS modules, the second group of feature extraction units includes downsampling and four Fire_DS modules, and the third group of feature extraction units includes downsampling and two Fire_DS modules.

[0042] As a further improvement to the method for classifying electroencephalogram signals of Parkinson's disease patients based on deep learning in the present invention:

[0043] The Fire_DS module is an improvement based on the fire module of the SqueezeNet network. The input features first go through a convolution operation and then are sent to a dual-branch structure. The left branch is a conventional convolution, and the right branch is a parallel depth convolution and channel attention module. Subsequently, the output features of the left branch and the right branch are concatenated along the channel dimension. In addition, an identity mapping mechanism is introduced, and the initial input features are directly added to the output features after concatenation of the left and right branches through a skip connection as the output features of the Fire_DS module.

[0044] As a further improvement to the method for classifying electroencephalogram signals of Parkinson's disease patients based on deep learning in the present invention:

[0045] The offline training process of the CA-ResFireNet network is as follows:

[0046] The training data comes from the publicly available Iowa dataset. Each sample in the Iowa dataset is subjected to the aforementioned preprocessing and the noise-assisted multi-dimensional empirical mode decomposition, and then the Shannon entropy feature matrix and the phase lag index feature matrix are extracted and fused into a Shannon entropy and phase lag index fusion matrix.

[0047] The data of the Shannon entropy and phase lag index fusion matrix is divided into a training set, a validation set, and a test set, and a label is established for each sample. During the offline training process, the CA-ResFireNet network is trained using the cross-entropy loss function and the AdamW optimizer.

[0048] The beneficial effects of the present invention are mainly reflected in:

[0049] First, it can improve time and frequency resolution: The present invention uses the noise-assisted multi-dimensional empirical mode decomposition technology to overcome the limitations of the Fourier transform and wavelet transform in processing non-stationary and complex EEG signals, and can achieve higher time and frequency resolution in the processing of non-stationary and complex EEG signals. For the Iowa dataset, the accuracy can reach 96.62%, proving that the present invention can capture the characteristics of EEG signals more accurately and is helpful for more effective detection of early disease characteristics.

[0050] Second, it can perform multi-level feature extraction: By extracting the Shannon entropy and phase lag index feature matrices, the present invention can not only capture time-frequency domain features, but also obtain the potential dependence correlations between brain regions of Parkinson's disease patients and the chaos and randomness of the EEG signals themselves, thus more comprehensively reflecting the rich information of the EEG signals, better mining the potential patterns in the data, and helping to improve the classification accuracy.

[0051] Finally, it can reduce the number of parameters and improve the calculation efficiency: Through the design of a lightweight neural network in the structural design of the present invention, the number of parameters and the calculation time are significantly reduced. While maintaining or even improving the classification accuracy, the model of the present invention is more suitable for deployment in the scenario of actual mobile devices, helping to improve the calculation speed and reduce the demand for computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The following further details the specific implementation manners of the present invention with reference to the drawings.

[0053] Figure 1 It is a schematic diagram of the EEG signal classification method for Parkinson's disease patients based on deep learning of the present invention;

[0054] Figure 2 It is a schematic diagram of the structure of the Fire_DS module of the present invention;

[0055] Figure 3 Schematic diagram for comparing the principles of depth convolution and standard convolution;

[0056] Figure 4 Schematic diagram of the depth residual structure;

[0057] Figure 5 Schematic diagram of the structure of the channel attention module;

[0058] Figure 6 Schematic diagram of the structure of the CA-ResFireNet network designed by the present invention. Detailed implementation manners

[0059] The present invention will be further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto:

[0060] Embodiment 1. A method for classifying electroencephalogram signals of Parkinson's disease patients based on deep learning, as Figure 1 shown. First, for high-precision analysis to improve time and frequency resolution, the present invention introduces a new analysis method, using noise-assisted multi-dimensional empirical mode decomposition technology to decompose electroencephalogram signal segments, so as to overcome the limitations of Fourier transform and wavelet transform in processing non-stationary and complex electroencephalogram signals, and can achieve higher resolution in the time and frequency domains, thereby more accurately capturing the characteristics of electroencephalogram signals. Secondly, by improving the Fire module of the SqueezeNet network and building a new CA-ResFireNet network, a multi-level feature extraction technology is introduced, which not only focuses on time-frequency domain features, but also includes other level features outside the time domain and frequency domain, so as to comprehensively capture the rich information of electroencephalogram signals and improve the classification accuracy; at the same time, in order to improve efficiency and speed up the calculation, the structure of the CA-ResFireNet network is lightened, reducing the number of parameters and calculation time, but maintaining or even improving the classification accuracy. This lightweight design will make the model more suitable for actual clinical applications, speed up the calculation, and reduce the computational resource requirements. Specifically:

[0061] Step 1: Collect electroencephalogram signals and perform preprocessing;

[0062] During the process of electroencephalogram (EEG) signal acquisition, due to the physiological activities of the subjects and external environmental interference, the collected EEG signals will inevitably contain various noises and artifacts. Suppressing and processing them are key steps to improve the quality of EEG signals and the accuracy of analysis. The preprocessing of EEG signals includes filtering, downsampling, artifact removal, and segmentation. First, band-pass filtering of 0.1 - 40 Hz is performed on the EEG signals of each channel to remove low-frequency and high-frequency noises. Then, in order to reduce the amount of data and improve the processing efficiency, the sampling rate of the EEG signals (256 HZ) is appropriately reduced to accelerate the subsequent processing rate. Next, independent component analysis technology is used to eliminate electrooculogram (EOG) artifacts caused by the eye movements of the subjects and electromyogram (EMG) artifacts caused by muscle tremors. The frequencies of these artifacts are close to those of EEG signals and will directly affect the accurate measurement of brain activities. Therefore, the removal of EOG and EMG artifacts is crucial. Finally, the EEG signals are segmented into non-overlapping time windows of 1 s to obtain the preprocessed EEG signal segments, which are convenient for subsequent feature extraction work.

[0063] Step 2: Use noise-assisted multi-dimensional empirical mode decomposition technology to perform multi-scale decomposition on the preprocessed EEG signal segments to obtain the intrinsic mode components of different frequency bands;

[0064] The core of the noise-assisted multi-dimensional empirical mode decomposition technology algorithm is to splice v-dimensional independent white noise into the s-dimensional input signal to form a signal to be decomposed with a dimension of s + v, and then use the multi-variate empirical mode decomposition method to decompose and process to obtain the corresponding intrinsic mode components. Discard the part of the intrinsic mode components corresponding to the white noise to obtain the intrinsic mode components of different frequency bands of the original signal. Let the preprocessed EEG signal segment be an EEG signal of a v-variable. The steps of the noise-assisted multi-dimensional empirical mode decomposition technology algorithm are as follows:

[0065] Step 2.1: Generate s-channel uncorrelated Gaussian white noise signals:

[0066]

[0067] where, n i (t), i = 1, 2, …, s are white noise signals, and T is the signal length;

[0068] Step 2.2: Splice the s-channel Gaussian white noise signals into the v-channel EEG signals to form an s + v-channel multi-variate signal:

[0069]

[0070] where, x i (t), i = 1, 2, …, v are v-channel EEG signals;

[0071] Step 2.4: Invoke the multivariate empirical mode decomposition algorithm to decompose the multivariate signal to generate M intrinsic mode components:

[0072]

[0073] where c m (t) is the intrinsic mode component corresponding to the input EEG signal a m (t) is the intrinsic mode component corresponding to the Gaussian white noise signal r M (t) is the remaining residue;

[0074] Step 2.4: Discard the intrinsic mode component a corresponding to the Gaussian white noise signal m to obtain the original input EEG signal and the intrinsic mode components IMF(t) in different frequency bands obtained after noise-assisted multi-dimensional empirical mode decomposition:

[0075]

[0076] Step 3: Extract the Shannon entropy feature matrix from the preprocessed EEG signal segments in Step 1, and extract the phase lag index feature matrix from the intrinsic mode components IMF(t) in different frequency bands after decomposition in Step 2;

[0077] The Shannon entropy can quantify the complexity and uncertainty of the EEG signal, is insensitive to noise, can effectively extract the hidden dynamic information in the EEG signal, improve the robustness of the model, is simple to calculate, and is suitable for real-time processing and rapid analysis. The phase lag index measures the phase synchronization between different brain regions, reflects the functional connection pattern between brain regions, is not affected by volume leads, can provide more accurate functional connection information, helps to avoid errors and improve classification accuracy.

[0078] Step 3.1: Extracting the Shannon entropy feature matrix includes the following steps:

[0079] (1) Calculate the Shannon entropy of the EEG signal of a single channel using the following formula:

[0080]

[0081] where p represents the probability of the one-dimensional EEG signal, represents the j-th sample of the EEG signal x i (t) of the i-th channel. For an EEG signal segment of size 63*250, calculate the Shannon entropy of each channel to obtain a feature vector ShEn s of size 63*1.

[0082] (2) Through the eigenvector ShEn s Construct a Shannon entropy feature matrix ShEn of size 63*63 f , and the calculation formula is as follows:

[0083] ShEn f (i,j) = ShEn s (i) + ShEn s ′(j) (6)

[0084] Among them, ShEn f (i,j) represents the element in the i-th row and j-th column of the Shannon entropy feature matrix, and ShEn s ′ is the transpose of ShEn s .

[0085] Step 3.2: Extract the phase lag index feature matrix

[0086] Take the first four of the intrinsic mode functions IMF(t) obtained from formula (4) to form a component matrix of size 4*63*250 (4 represents the number of components in different frequency bands, 63 is the number of EEG signal channels, and 250 is the number of sampled data points). For each 63*250-sized component matrix, calculate the phase lag index between 63 EEG channels in turn. The formula is as follows:

[0087]

[0088] Among them, N represents the time point, and Φ rel represents the phase difference between the signals of two channels at time t n . sign is a sign function. When the independent variable is positive, its output is 1. When the independent variable is negative, the result is -1. For 0, the result is also 0. After calculating the phase lag index for the four component matrices, a phase lag index feature matrix of size 4*63*63 can be obtained.

[0089] Then, fuse the Shannon entropy feature matrix and the phase lag index feature matrix to obtain a Shannon entropy and phase lag index fusion matrix of size 5*63*63, which is used as the input of the CA-ResFireNet network of the present invention.

[0090] Step 4: Build a neural network model. While introducing the depth convolution operator and the residual structure, combine the channel attention mechanism to extract the deep features of Parkinson's disease and obtain the classification result of the EEG signal.

[0091] The traditional convolutional neural network SqueezeNet architecture is composed of stacked basic modules, the fire module, which contains two parts: the squeeze layer and the expand layer. The squeeze layer uses a 1*1 convolutional kernel for feature compression to reduce the dimension of the feature vector, while the expand layer uses 1*1 and 3*3 convolutional kernels respectively for feature dimension expansion, and finally performs concat. It aims to reduce the number of parameters and computational complexity through efficient convolutional operations and model compression, but it has deficiencies in feature extraction ability and information transmission efficiency.

[0092] Based on the fire module, the Fire_DS module is designed and built in the present invention. The 1*1 convolutional kernel of the original left branch of the expand layer is replaced with a 3*3 convolutional kernel, and the 3*3 convolution of the original right branch of the expand layer is replaced with a 3*3 depthwise convolution and a parallel channel attention module, as Figure 2 shown. The purpose of replacing the original 3*3 convolution with a 3*3 depthwise convolution is to improve the feature extraction ability of the model as much as possible while the number of parameters is within an acceptable range, and the channel attention module assigns different weights to the channels, enabling the model to focus on important channels and ignore irrelevant features, thereby improving the efficiency of model feature extraction. The output of the channel attention module is a one-dimensional vector, with a size between 0 and 1, representing the weight coefficient of each channel. The higher or lower the coefficient indicates the importance of the corresponding channel data. Therefore, the output of the depthwise convolution is multiplied by the output of the attention module to achieve that the features of important channels account for a large proportion and the features of unimportant channels account for a small proportion.

[0093] The output of the 3*3 depthwise convolution is multiplied by the output of the channel attention module, and then concatenated with the output features of the left branch of the expand layer. Finally, combining the idea of the deep residual network, an identity mapping branch is added to the module to prevent overfitting as the depth of the model increases.

[0094] (1) For the depthwise convolution operator, it convolves each input feature channel separately compared to ordinary convolution operations. Therefore, the number of convolutional kernels is equal to the number of input feature channels and also equal to the number of output feature maps. Depthwise convolution can greatly reduce the number of parameters and computational amount. The principle comparison between it and standard convolution is shown in Figure 3 .

[0095] Assume the convolutional kernel size is Dk*Dk, the input channels are M, the output channels are N, and the output feature map size is DF*DF. After standard convolution, it can be calculated that:

[0096] Number of parameters: DK*DK*M*N

[0097] Computational complexity: DK * DK * M * N * DF * DF

[0098] When using depthwise convolution operation, the convolution kernel size is DK * DK * 1 and the number of convolution kernels is M. Therefore, it can be calculated as follows:

[0099] Number of parameters: DK * DK * M

[0100] Computational complexity: DK * DK * M * DF * DF

[0101] By comparing standard convolution and depthwise convolution, it can be found that both the computational complexity and the number of parameters of depthwise convolution are 1 / N of those of standard convolution. Therefore, by combining depthwise convolution operators, the number of parameters and the computational complexity can be extremely effectively reduced, and the computational efficiency of the network can be improved.

[0102] (2) For the deep residual structure, the core idea is to add a new identity mapping branch in parallel in the network module, transfer the input features to the output layer, and add them to the features generated by the module, so as to retain the original feature information. The structure is shown in Figure 4 .

[0103] The calculation formula is as follows:

[0104] y = F(x, W i ) + x (8)

[0105] where F(x, W i ) is the output of the layer with weight W i , x is the input feature vector, and y is the output feature vector.

[0106] In a deep neural network, the vanishing gradient will cause the model to be difficult to converge and affect the training effect. The residual architecture adds an identity mapping in the backbone network, enabling the gradient to propagate more effectively in the network, thus alleviating the problem of vanishing gradient. At the same time, the deep residual structure can help the model better learn deep and complex features, reduce the risk of overfitting, and improve the generalization ability of the model.

[0107] (3) For the channel attention module, the structure schematic diagram is shown in Figure 5 , and its function is described as follows:

[0108] The main function of the channel attention module is to enhance the network's ability to adaptively assign weights to the importance of different channel features. First, the input feature map is globally average pooled in the spatial dimension, and the feature map u of W×H×C containing global information is directly compressed into a feature vector Z of 1×1×C. This step is equivalent to aggregating the global information of each channel to extract the global features at the channel level. The formula is as follows:

[0109]

[0110] Among them, Z c is the c-th element of the feature vector Z, H and W are the height and width of the feature map respectively, and u c (i, j) is the element at the i-th row and j-th column of the input feature map on the c-th channel.

[0111] After that, a gate mechanism composed of two fully connected layers is used to comprehensively capture channel dependencies. Specifically, the first fully connected layer compresses C channels into C / r channels to reduce the computational complexity, and then passes through a RELU non-linear activation layer. The second fully connected layer restores the number of channels back to C channels to obtain the weight s of size 1×1×C. Finally, the elements of the weight s are mapped between 0 and 1 through the sigmoid function, where r is the compression ratio. The formula is as follows:

[0112] s = σ(W2δ(W1z)) (10)

[0113] δ = max(0, 1) (11)

[0114] Among them, δ represents the RELU activation function, σ represents the sigmoid function, and W1 and W2 are the weight matrices of the corresponding fully connected layers.

[0115] Finally, the attention weights are weighted to the features of each channel to obtain the final output X′, that is, multiplying it with the input feature u c The formula is as follows:

[0116] X′ = s c ×u c (12)

[0117] Among them, s c is the coefficient corresponding to the c-th channel of the weight matrix s, and u c is the element corresponding to the c-th channel of the input feature u.

[0118] With the embedding of the channel attention module, the Fire_DS module can enhance the weights of important feature channels, suppress unimportant feature channels, and improve the representation ability of the model.

[0119] In the present invention, the fire_DS module is obtained by optimizing and improving the fire module. The CA-ResFireNet network is constructed by repeatedly stacking the fire_DS module and ending with a global pooling layer and a softmax layer. The CA-ResFireNet model of the present invention is as Figure 6As shown, the Shannon entropy and phase lag index fusion matrix with a size of 5*63*63 obtained in step 3 first passes through 3 groups of feature extraction units composed of downsampling and Fire_DS modules in sequence (where the first group of feature extraction units is downsampling and three Fire_DS modules, the second group of feature extraction units is downsampling and four Fire_DS modules, and the third group of feature extraction units is downsampling and two Fire_DS modules), gradually extracting the functional connectivity features and Shannon entropy characteristics of brain regions related to Parkinson's disease. During this process, through the synergistic effect of depth convolution, channel attention mechanism, and residual structure, the expression ability of key features is enhanced. Subsequently, the feature map is processed by the global pooling layer to integrate the high-dimensional features into a one-dimensional vector, effectively retaining the most representative global feature information. Finally, the classification result is normalized by the Softmax layer to output the probability distribution corresponding to each category, thereby realizing the classification prediction of EEG signals of Parkinson's disease.

[0120] Step Five: Offline training of the neural network.

[0121] The training data of the present invention comes from the Iowa dataset publicly available from the University of Iowa, which records their EEG data in the resting state with eyes open, including 14 Parkinson's disease patients and 14 healthy controls, a total of 28 subjects. During the signal acquisition process, a 64-channel Brain Vision system was used for acquisition, with a sampling rate of 500 hz and a sampling duration of 2 minutes.

[0122] Each sample in the dataset is preprocessed through step 1, and then, the noise-assisted multi-dimensional empirical mode decomposition technique is used to perform multi-scale decomposition processing on the preprocessed EEG signal segments, and then the phase lag index feature matrix is extracted; in addition, the Shannon entropy feature matrix is extracted from the preprocessed EEG signal segments, and the Shannon entropy feature matrix and the phase lag index feature matrix are fused into a Shannon entropy and phase lag index fusion matrix, which is randomly divided into a training set, a validation set, and a test set (with a ratio of 8:1:1). Then, each sample in the training set, validation set, and test set is manually labeled with labels, using 0 and 1 to represent the EEG signals of Parkinson's disease patients and healthy individuals respectively. The training set is used for the learning and training of the model to adjust the weight parameters to make the model fit the training data as much as possible; the validation set is used for the tuning and selection of the model to evaluate the model effect; the test set is used to evaluate the generalization ability of the model on a new dataset and test the classification performance of the model. During the training process, the cross-entropy loss function and the AdamW optimizer are used, the learning rate is set to 0.0001, and the Batchsize is set to 32.

[0123] Step Six: Online use

[0124] Load the offline-trained CA-ResFireNet network model into the cloud server and deploy the deep learning inference framework. Then, transmit the obtained electroencephalogram (EEG) signals of the patient to the system for preprocessing, including downsampling, filtering and denoising, artifact removal, and feature extraction. Immediately afterwards, perform real-time inference. For the preprocessed EEG signal segments, use the noise-assisted multi-dimensional empirical mode decomposition technique for multi-scale decomposition processing to obtain the intrinsic mode components in different frequency bands, and then extract the phase lag index feature matrix; in addition, extract the Shannon entropy feature matrix for the preprocessed EEG signal segments, and then fuse the two into the Shannon entropy and phase lag index fusion matrix; then input the fused 5*63*63 Shannon entropy and phase lag index fusion matrix into the deployed CA-ResFireNet network. The network extracts the key classification features through forward propagation and outputs the probability distributions of various categories.

[0125] Experiment:

[0126] The experimental data comes from the publicly available Iowa dataset of the University of Iowa.

[0127] 1. Comparative Experiment 1

[0128] Conduct an experimental comparison between the noise-assisted empirical mode decomposition technique adopted in the present invention and the Fourier transform and wavelet transform methods to verify the advantages of the method of this patent in time-frequency resolution.

[0129] In this experiment, the Iowa dataset is uniformly used. The dataset is preprocessed to remove artifacts and noise interference. The Shannon entropy feature matrix is extracted from the preprocessed data. At the same time, the preprocessed data is respectively processed as follows:

[0130] Method 1 (the present invention): Decompose through the noise-assisted empirical mode decomposition technique of the present invention, extract the phase lag index feature matrix, then fuse the two into the Shannon entropy and phase lag index fusion matrix, classify through the CA-ResFireNet network of the present invention, and count the accuracy rate;

[0131] Method 2: Decompose using the Fourier transform, extract the phase lag index feature matrix, then fuse the two into the Shannon entropy and phase lag index fusion matrix, classify through the CA-ResFireNet network of the present invention, and count the accuracy rate;

[0132] Method 3: Decompose using the wavelet transform, extract the phase lag index feature matrix, then fuse the two into the Shannon entropy and phase lag index fusion matrix, classify through the CA-ResFireNet network of the present invention, and count the accuracy rate.

[0133] Method 1 achieved an accuracy of 96.62%, Method 2 obtained an accuracy of 96.25%, and Method 3 achieved an accuracy of 96.07%. The results show that the noise-assisted multi-dimensional empirical mode decomposition technology of the present invention can provide higher-resolution time-frequency features and improve the performance of model classification.

[0134] 2. Comparative Experiment II

[0135] Feature matrices were constructed using different feature generation methods respectively. By analyzing the influence of the feature matrix generation methods proposed in different literatures on the classification performance, including calculating the differential entropy feature matrix 0 and the power spectral density feature matrix 0, the advantages of the feature matrix generation method in this patent were evaluated.

[0136] In this experiment, the Iowa dataset was uniformly used. First, the dataset was preprocessed, and the preprocessed data was respectively:

[0137] Method 4: According to the literature 【1】 Calculate the differential entropy feature matrix; classify and detect electroencephalogram signals through the CA-ResFireNet network designed by the present invention, and count the accuracy rate.

[0138] Method 5: According to the literature 【2】 Calculate the power spectral density feature matrix; classify and detect electroencephalogram signals through the CA-ResFireNet network designed by the present invention, and count the accuracy rate.

[0139] The experimental results show that the classification accuracy rate of Method 4 is 92.24%, and the classification accuracy rate of Method 5 is 90.43%, both of which are lower than the accuracy rate of 96.62% achieved by Method 1 of the present invention. Therefore, the method proposed in this patent can provide key classification information for the network and significantly improve the classification accuracy rate.

[0140] 3. Comparative Experiment III

[0141] Compare the performance of the ResNet 【3】 , MobileNet 【4】 , MobileVit 【5】 classification networks with the CA-ResFireNet network designed by the present invention to verify the effectiveness and advancement of the proposed network structure. The test set constructed in step 5 of Example 1 was uniformly used in Comparative Experiment III.

[0142] Method 6: Use the ResNet network to classify and detect electroencephalogram signals, and the counted accuracy rate is 99.52%;

[0143] Method 7: Use the MobileNet network to classify and detect electroencephalogram signals, and the counted accuracy rate is 99.44%;

[0144] Method 8: Use the MobileVit network for the classification and detection of electroencephalogram signals, and the statistical accuracy rate is 95.59%;

[0145] Method 9: Use the CA-ResFireNet network designed by the present invention for the classification and detection of electroencephalogram signals, and the statistical accuracy rate is 96.62%.

[0146] Finally, the model of the present invention achieved the best classification effect, obtaining an accuracy rate of 96.62%. Compared with MobileVit, which is the best-performing in the comparison network, the accuracy rate was increased by 1.03%. In addition, in terms of the number of parameters, the number of parameters of our model is only 0.85M. In contrast, the number of parameters of ResNet, MobileNet, and MobileVit are 21.8M, 2.23M, and 0.95M respectively. The results show that the network proposed in this patent is superior to the existing networks in terms of classification accuracy, the number of parameters, and computational efficiency.

[0147] The references involved in the above text are as follows:

[0148] 【1】D. Shah, G. K. Gopika, N. Sinha, Analysis of eeg for parkinson’s diseasedetection, in: 2022 IEEE International Conference on Signal Processing andCommunications (SPCOM), IEEE, 2022, pp. 1–5;

[0149] 【2】S. Wang, G. Wang, G. Pei, T. Yan, An eeg-based approach for parkinson’sdisease diagnosis using capsule network, in: 2022 7th International Conferenceon Intel-ligent Computing and Signal Processing (ICSP), IEEE, 2022, pp. 1641–1645;

[0150] 【3】K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778;

[0151] 【4】A. Howard, A. Zhmoginov, L.-C. Chen, M. Sandler, M. Zhu, Inverted residuals and linear bottlenecks: Mobile networks for classification, detection and segmentation, in: Proc. CVPR, 2018, pp. 4510–4520;

[0152] 【5】S. Mehta, M. Rastegari, Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer, arXiv preprint arXiv:2110.02178.

[0153] Finally, it should also be noted that the above-listed are only several specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and there can be many variations. All variations that can be directly derived or associated by those of ordinary skill in the art from the disclosed content of the present invention shall be considered as within the protection scope of the present invention.

Claims

1. A deep learning-based method for classifying EEG signals of Parkinson's disease patients, characterized by: S1, collecting EEG signals of patients with Parkinson's disease, and then preprocessing to obtain EEG signal fragments; S2, subjecting the EEG signal segments to noise-assisted multidimensional empirical mode decomposition to obtain intrinsic mode components of different frequency bands, and then extracting a phase lag index feature matrix; S3, extracting the Shannon entropy feature matrix from the EEG signal segment, and then fusing the Shannon entropy feature matrix and the phase lag index feature matrix into a Shannon entropy and phase lag index fusion matrix; S4, Shannon entropy and phase lag index fusion matrix are used to obtain the classification results and confidence of EEG signals through the CA-ResFireNet network trained offline.

2. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 1, characterized in that: The preprocessing includes bandpass filtering, down sampling, artifact removal and segmentation processing.

3. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 2, characterized in that: The noise-assisted multidimensional empirical mode decomposition is specifically: (1) Generate Gaussian white noise signal: Among them, n i (t),i=1,2,L,s is the white noise signal, T is the signal length; (2) Concatenate the Gaussian white noise signal and the EEG signal fragment into a multivariate signal: Among them, x i (t), i = 1, 2, L, v is the EEG signal of channel v, representing the EEG signal segment; (3) Multivariate empirical mode decomposition algorithm is used to analyze multivariate signals Decompose and generate M intrinsic modal components Among them, c m (t) is the intrinsic modal component corresponding to the EEG signal segment, a m (t) is the intrinsic modal component corresponding to the Gaussian white noise signal, r M (t) is the remaining residue; (4) Discard the intrinsic modal component a in formula (3) m (t): Wherein, IMF(t) represents the intrinsic modal components of different frequency bands obtained after the EEG signal segment is subjected to noise-assisted multidimensional empirical mode decomposition.

4. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 3, characterized in that: The process of extracting the Shannon entropy feature matrix is ​​as follows: (1) Calculate the Shannon entropy of a single-channel EEG signal: Among them, p represents the probability of a one-dimensional EEG signal, Represents the EEG signal x of the i-th channel i The jth sample of (t); (2) Construct the Shannon entropy feature matrix ShEn f : ￠ ShEn f (i,j)=ShEn s (i)+ShEn s (j) (6) Among them, ShEn f (i,j) represents the element in the i-th row and j-th column of the Shannon entropy feature matrix. s (i) represents the Shannon entropy of the i-th channel, ShEn s ￠ is ShEn s The transpose of .

5. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 4, characterized in that: The process of extracting the phase hysteresis index characteristic matrix is ​​as follows: Take the first four components of the intrinsic modal component IMF(t) to form a feature matrix of size 4*63*250. For each component matrix of size 63*250, calculate the phase lag index between the EEG channels in turn. The formula is as follows: Where N represents the time point, F rel Represents the two channel signals at time t n The phase difference at , sign is the sign function; After calculating the phase lag index of the four component matrices, a phase lag index characteristic matrix of 4*63*63 size is obtained.

6. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 5, characterized in that: The CA-ResFireNet network is improved based on the SqueezeNet network, including three groups of feature extraction units, a global pooling layer and a Softmax layer. The first group of feature extraction units includes downsampling and three Fire_DS modules, the second group of feature extraction units includes downsampling and four Fire_DS modules, and the third group of feature extraction units includes downsampling and two Fire_DS modules.

7. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 6, characterized in that: The Fire_DS module is an improvement on the fire module based on the SqueezeNet network. The input features are first convolved and then sent to a dual-branch structure. The left branch is a conventional convolution, and the right branch is a parallel deep convolution and channel attention module. Subsequently, the output features of the left and right branches are concatenated along the channel dimension. In addition, an identity mapping mechanism is introduced, and the initial input features are directly added to the output features of the left and right branches through jump connections as the output features of the Fire_DS module.

8. The method for classifying EEG signals of Parkinson's disease patients based on deep learning according to claim 7, characterized in that: The offline training process of the CA-ResFireNet network is as follows: The training data comes from the public Iowa data set. Each sample in the Iowa data set is subjected to the preprocessing and the noise-assisted multidimensional empirical mode decomposition, and then the Shannon entropy feature matrix and the hysteresis index feature matrix are extracted and fused into a Shannon entropy and hysteresis index fusion matrix. The data of the Shannon entropy and hysteresis index fusion matrix are divided into training set, validation set and test set, and each sample is labeled; The CA-ResFireNet network is trained using the cross entropy loss function and the AdamW optimizer during offline training.

Images