A Parkinson's Disease Identification Method Based on Frequency Domain and Spatiotemporal Features of EEG Signals

Abstract

This invention discloses a method for identifying Parkinson's disease based on frequency domain and spatiotemporal features of electroencephalogram (EEG) signals. The method first extracts the average power of the Delta, Theta, and Alpha frequency bands from multi-channel EEG signals, and concatenates the average power features of the three frequency bands into a one-dimensional vector as a feature matrix, which is then fed into an SVM classifier for classification. Next, a compact convolutional neural network based on spatiotemporal features is constructed to iteratively extract the spatiotemporal features of the EEG signals, while simultaneously introducing a channel attention mechanism to adaptively correct the output feature channels. Finally, the optimal weights of the two models are searched within the [0,1] interval, and the two models are fused according to the weights to obtain the final classification result. This method has broad application prospects in the field of EEG signal recognition and Parkinson's disease classification.

Description

Technical Field

[0001] This invention designs a method for identifying Parkinson's disease based on frequency domain and spatiotemporal features of electroencephalogram (EEG) signals, which relates to the field of EEG signal classification technology. Background Technology

[0002] Parkinson's disease (PD) is the most common motor disorder, a chronic, degenerative, and progressive disease, and the second most common neurodegenerative disease after Alzheimer's disease (AD). The average age of onset for PD is over 60 years old; however, in rare cases, symptoms may appear earlier in younger people. Neurodegenerative diseases are a leading cause of declining brain function and quality of life in older adults, affecting 1-2% of those over 65. In later stages of Parkinson's disease, patients experience various motor and non-motor complications. Early detection and accurate diagnosis of PD symptoms are crucial to combating these complications. To this day, the origin of PD remains unknown, and due to the lack of treatment options for Parkinson's disease, research into early PD detection methods is essential to slow disease progression and improve patients' quality of life.

[0003] Although various physiological and pathological findings exist in the diagnosis of Parkinson's disease (PD), these findings do not provide complete evidence to distinguish between healthy and diseased individuals. Diagnosis is challenging when no obvious motor or non-motor symptoms are observed. Currently, there are no definitive single biomarkers, such as blood biomarkers, that can accurately detect early-stage PD, nor can brain scans or other tests definitively diagnose Parkinson's disease. The most effective approach remains clinical diagnosis by physicians based on the patient's medical history, behavior, and cross-examination in the initial stages of PD or when levodopa efficacy is low, followed by reassessment of the patient's physical and neurological symptoms through follow-up. However, obvious motor symptoms may only be present when most dopaminergic neurons have been lost, making diagnosis difficult when individuals lack obvious signs or symptoms. Early prodromal symptoms play a crucial role in the diagnosis of early-stage PD, as by the time motor symptoms appear, 70% of dopaminergic synapses have already died, making it too late even with neuroprotective agents.

[0004] Parkinson's disease has become an increasingly serious threat, significantly impacting patients' quality of life. Reassessing physical and neurological symptoms in the early stages of Parkinson's disease or when levodopa's effects are weak using clinical characteristics is a subjective process. Therefore, there is an urgent need for an objective diagnostic method to support physicians in the diagnosis and treatment of Parkinson's patients and to detect the initial stages of the disease. Electroencephalography (EEG) signals, due to their low spatial sensitivity and high temporal resolution, are used for functional analysis of brain activity. Analysis of EEG signals can yield rich information about brain function, which is crucial for the detection and diagnosis of various types of neurological disorders. Summary of the Invention

[0005] This invention proposes a method for identifying Parkinson's disease based on frequency domain and spatiotemporal features of electroencephalogram (EEG) signals for classification tasks. The method first extracts the average power of the Delta, Theta, and Alpha frequency bands from multi-channel EEG signals and concatenates these three sub-bands into a one-dimensional vector as the frequency domain feature of the EEG, which is then fed into an SVM classifier for classification. Next, a compact convolutional neural network based on spatiotemporal features is constructed to iteratively extract the spatiotemporal features of the EEG signals, while simultaneously introducing a channel attention mechanism to adaptively correct the output feature channels. Finally, the optimal weights for the two models are searched within the [0,1] interval, and the two models are fused according to these weights to obtain the final classification result.

[0006] The present invention achieves the above objectives through the following technical solutions:

[0007] A method for identifying Parkinson's disease based on electroencephalogram (EEG) signals with frequency domain and spatiotemporal characteristics includes the following steps:

[0008] Step 1: First, preprocess the raw EEG signal by filtering, artifact removal, and data augmentation.

[0009] Step 2: Use the Fast Fourier Transform (FFT) method to obtain the average power of the Delta, Theta, and Alpha frequency bands from the multi-channel EEG as feature matrices, and then feed them into the classifier for training.

[0010] Step 3: Construct a compact convolutional neural network based on spatiotemporal features and introduce a channel attention mechanism to adaptively extract effective features representing Parkinson's disease;

[0011] Step 4: Finally, by searching within the [0,1] interval with a width of 0.05, the weight values ​​of the two models are calculated, and the two models are fused based on the weights to obtain the final classification result.

[0012] The role of the preprocessing operation of the raw EEG signal in step one is explained as follows:

[0013] Preprocessing of EEG signals is crucial for extracting valuable information. This preprocessing includes filtering, artifact removal, and data augmentation. Since noise is introduced during signal acquisition due to various external influences, filtering out noise interference and retaining useful frequency bands before analysis is essential. A fourth-order Butterworth bandpass filter was used to filter the EEG signals, with a cutoff frequency set to 0.5-12Hz, while simultaneously filtering out the 60Hz power frequency. Artifacts in EEG signals mainly include those caused by eye movements, ECG, and EMG. Therefore, effective artifact removal is vital for signal analysis. Independent principal component analysis (IPA) based on blind source separation was used to decompose the signal into multiple independent source signals, removing artifacts caused by eye movements, ECG, and EMG. Since the sample size of the dataset is relatively small, in order to expand the data volume, the signal is first divided into short sequences of 1 second in terms of no overlap, 30% overlap, and 50% overlap in time. Then, during the training phase, a random number is generated when each data segment is input. If the random number is greater than 0.5, random noise is added to the data; if the random number is less than 0.5, no noise is added.

[0014] The role of the attention module in step three is explained below:

[0015] By adaptively recalibrating feature maps by embedding channel attention modules into a pre-constructed compact convolutional neural network based on spatiotemporal features, meaningful features are enhanced while weaker features are suppressed. Since the feature matrix on each channel represents the components represented by different convolutional kernels, the feature information is decomposed into information on C convolutional kernels. These C information components contain all the key information points of the original EEG signal. By assigning different weights to these C channels, the correlation between different channel components and key information in the EEG signal can be represented. Calculating these correlations allows attention to key parts of the image EEG signal, achieving the goal of eliminating useless information.

[0016] Global average pooling is used to operate on the input feature map X = [x1, x2, ..., xn], where x i ∈R H×W Representing each channel, a 1×1×C vector denoted as z is generated. This involves summing the values ​​of each channel and averaging them to integrate the feature map from a spatial perspective. This vector contains the global distribution of the channel feature responses, enabling the acquisition of information about the global receptive field. As shown in formula (1), H and W represent the height and width of the input feature map, respectively, and z... m The m-th value in the compressed z-vector, x m (i,j) is the value of the m-th layer, i-th row and j-th column of the input feature map.

[0017]

[0018] Then, two fully connected layers are used to encode the z vector to obtain the importance of each channel. After the first fully connected layer, the ReLU activation function is used to perform a nonlinear operation on the vector. As shown in Equation (2), W1 and W2 are the weights of the first and second fully connected layers, respectively, and z′ is the vector after the nonlinear operation on z.

[0019] z'=W2(ReLU(W1z)) (2)

[0020] Finally, a sigmoid function is used to map the channel importance to the [0,1] interval. The generated weight vector is then multiplied by the original feature map to recalibrate it. The reweighted feature map is then used as input for subsequent layers in training. As shown in Equation (3), X' is the reweighted feature map, σ is the sigmoid operation, and z n ' is the nth value in the vector z'.

[0021] X'=[x1σ(z1'),x2σ(z2'),···,x n σ(z n ')] (3) Attached Figure Description

[0022] Figure 1 This is the overall framework of the present invention.

[0023] Figure 2 This invention is a compact convolutional neural network framework based on spatiotemporal features.

[0024] Figure 3 This is the channel attention module structure of the present invention. Detailed Implementation

[0025] Table 1 shows the classification performance of this invention, which integrates the average power of the Delta, Theta, and Alpha frequency bands as a feature matrix, under multiple classifiers.

[0026] Table 2 compares the results of introducing the channel attention module into the compact convolutional neural network with those of the basic network and frequency domain feature classification.

[0027] Table 3 shows the comparison results between the attention network based on frequency domain and spatiotemporal features proposed in this invention and different algorithms.

[0028] Table 1 shows the performance of fusing the average power features of the three frequency bands across multiple classifiers.

[0029]

[0030] A comprehensive analysis of the data in Table 1 shows that using the average power of the Delta, Theta, and Alpha bands can characterize the frequency domain features of EEG signals to a certain extent and can effectively distinguish between Parkinson's patients and normal subjects. Among the four classifiers, the SVM classifier exhibits the best classification performance, achieving a sensitivity of 74.46%, a specificity of 75.02%, and an accuracy of 74.74%. The results indicate that the extracted frequency domain feature matrix can effectively capture and characterize the relationships between characteristic attributes of the input signal, such as relative position, scale, and orientation. Therefore, this feature matrix is ​​used, and the SVM classifier is selected for classification.

[0031] Table 2 Comparison of Frequency Domain Features, STCN, and STACN Classification Results

[0032]

[0033] Experiments were conducted on the frequency domain feature model, the un-attentional convolutional neural network model (STCN), and the attention-added convolutional neural network model (STACN), and the classification results are shown in Table 2. The TPR index indicates whether there are missed diagnoses of Parkinson's disease patients; the higher the value, the stronger the model's detection ability. Table 2 shows that after introducing the channel attention module, the TPR improved by 10.34% compared to the frequency domain feature model and by 1.55% compared to STCN. Simultaneously, TNR and Acc were both higher than both the frequency domain feature model and STCN. This indicates that the model with channel attention has improved its detection ability for PD patients.

[0034] Table 3 Comparison of STFACN results with different algorithms

[0035]

[0036] The frequency domain feature model and the STACN model were fused, and the classification results of the fused network (STFACN) were compared with other methods. The comparison results are shown in Table 3. As can be seen from the table, the fused model STFACN shows a certain degree of improvement compared to both the frequency domain feature model and the STACN model, indicating that fusing frequency domain features and spatiotemporal features can better represent the EEG signals of Parkinson's patients. Compared with the methods mentioned above, the accuracy of the STFACN model is improved by 10.7%, 21.8%, 14.4%, and 3.1% respectively compared to the methods proposed by Chaturvedi, Yuvaraj, Vanneste, and Khoshnevis. The results show that this model can learn effective PD EEG feature representations, learn from 1 second of input data, and automatically extract deep features from EEG signals.

Claims

1. A method for identifying Parkinson's disease based on frequency domain and spatiotemporal features of electroencephalogram (EEG) signals, characterized in that... Includes the following steps: Step 1: First, preprocess the raw EEG signal by filtering, artifact removal, and data augmentation. Step 2: Calculate the average power of the Delta, Theta, and Alpha bands from the multi-channel EEG using Fast Fourier Transform (FFT). Then, calculate the energy of different sub-bands in each channel using FFT, and finally calculate the average power of each sub-band based on the energy. The formula is shown below: （1） （2） （3） Indicates the number of sampling points in the sample. Represents the original signal. This indicates the Fast Fourier Transform operation. The signal after Fourier transform. Represents the energy of the signal. express transpose, The average power of each frequency band is represented by the vector. Finally, the average power values ​​of all sub-bands are concatenated into a one-dimensional vector, which serves as the frequency domain feature of the EEG. The formula is shown below: （4） This represents the concatenated feature matrix. This represents a splicing operation. Indicates the first Each sample is processed, and finally the feature matrix is ​​fed into the classifier for classification. Step 3: Construct a compact convolutional neural network based on spatiotemporal features and introduce a channel attention mechanism to adaptively extract effective features representing Parkinson's disease; Step 4: Finally, by searching within the [0,1] interval with a width of 0.05, the weight values ​​of the two models are calculated, and the two models are fused based on the weights to obtain the final classification result.

2. The method for identifying Parkinson's disease based on frequency domain and spatiotemporal features of EEG signals according to claim 1, characterized in that, The preprocessing operation in step one involves first filtering the EEG signal using a fourth-order Butterworth bandpass filter with a cutoff frequency set to 0.5-12Hz, while simultaneously filtering out the 60Hz power frequency. Next, independent principal component analysis (IPA) based on blind source separation is used to decompose the signal into multiple independent source signals, removing artifacts caused by eye movements, ECG, and EMG. Then, the signal is divided into 1-second short sequences using non-overlapping, 30% overlap, and 50% overlap methods. During the training phase, a random number is generated simultaneously with each data segment input. If the random number is greater than 0.5, random noise is added to the data; if the random number is less than 0.5, no noise is added, thus expanding the training samples.

3. The method for identifying Parkinson's disease based on frequency domain and spatiotemporal features of EEG signals according to claim 1, characterized in that, Step three involves constructing a compact convolutional neural network based on spatiotemporal features. This involves directly designing a low-complexity, low-parameter four-layer compact CNN architecture to iteratively aggregate information from the spatial and temporal dimensions of the EEG signal. The model extracts spatial features in layers 1 and 3, and temporal features in layer 2, representing the EEG signal in a compact manner. The model's input is a preprocessed 1-second EEG signal with an input size of [insert size here]. In the first layer, 16 elements of size were learned. The convolutional kernel is used to extract the spatial features of the EEG signal at each time point; then... The activation function yields a nonlinear mapping result. The calculation formula is shown below; （5） Then, normalization is performed along the feature map dimension, and dropout is set to 0.25 to reduce network parameters; the output is then transposed to generate a new two-dimensional time series, and finally, data imputation is performed. The second layer first learns four features of size [value missing] based on the output features of the first layer. Two-dimensional convolution kernels, then using Max pooling is used to reduce dimensionality; the third layer learns four values ​​of size 1. The two-dimensional convolution kernel also uses Max pooling achieves feature dimensionality reduction; in the second and third layers... The activation function, normalization, dropout, and data imputation operations are the same as in the first layer; in the fourth layer, the features output from the third layer are first averaged and then fed into... Its output passes through The function yields probability values ​​for two classes; cross-entropy is used as the loss function to distinguish between the Parkinson's disease and normal categories, as shown in the formula below; （6） in, This represents the probability of the model's expected output. This represents the probability of the model's actual output. The cross-entropy loss function represents the model; A channel attention module is introduced, which assigns different weights to each channel to represent the correlation between different channel components and key information of the EEG signal; Global average pooling is used to map the input features. The operation involves summing all the values ​​from each channel, averaging them, and integrating the feature maps from a spatial perspective to generate a... The vector is denoted as ,in The formulas for each channel are shown below; （7） , These represent the height and width of the input feature map, respectively. After compression The first in the vector One value, It is the first of the input feature maps Layer, number Line number The value of the column; Then use two fully connected layers to... The vectors are encoded to obtain the importance of each channel, and then used after the first fully connected layer. The activation function performs a non-linear operation on the vector, as shown in the formula below; （8） , These are the weights of the first and second fully connected layers, respectively. To The vector after nonlinear operation; Finally passed The function maps channel importance to the interval [0,1], multiplies the generated weight vector with the original feature map, and uses the reweighted feature map as input to subsequent layers for further training, as shown in the formula; （9） For reweighted feature maps, for operate, for The first in the vector Values.

Images