Neural stress injury diagnosis method based on electroencephalogram data channel selection and deep learning neural network

Abstract

The invention discloses a neural stress injury diagnosis method based on electroencephalogram data channel selection and a deep learning neural network, and relates to the technical field of neural stress injury diagnosis, comprising the following steps: S1, a data preprocessing module; s2, an adaptive channel selection module; and S3, a feature classification module and a classification module. According to the neural stress injury diagnosis method based on electroencephalogram data channel selection and the deep learning neural network, EEG signals are converted into time-frequency diagrams through short-time Fourier transform (STFT), frequency characteristics changing along with time are effectively captured, signal characterization is enriched, and the accuracy of stress injury diagnosis is improved. By introducing an attention mechanism, key EEG channels related to stress injury are automatically identified and focused, personalized channel selection for different individuals is realized, and the diagnosis precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of neural stress injury diagnosis, and specifically to a neural stress injury diagnosis method based on EEG data channel selection and deep learning neural network. Background Art

[0002] Stress-induced injury refers to physical or psychological damage caused by external pressure or stress factors, which seriously affects the patient's life. In recent years, electroencephalogram (EEG) technology has become an important tool for stress injury diagnosis research. EEG can record and analyze the brain's electrical activity in real time, helping researchers monitor and evaluate changes in brain function under stress. With the rapid development of deep learning technology, EEG-based stress injury diagnosis models have performed well. Researchers use deep learning models to automatically extract and learn complex features in EEG signals, improving the efficiency and reliability of early diagnosis of stress injuries. These advances have significantly enhanced the ability to identify the stress state of soldiers and provided strong support for timely intervention and response.

[0003] There are currently no patents related to stress diagnosis based on EEG signals, but there are similar patents based on EEG for detecting neurological diseases: "A psychological stress detection method based on EEG signals" applied by Beijing Brain Land Technology Co., Ltd. (application number: CN202010822827.6) and "Feature extraction and automatic identification method for epilepsy EEG signals" applied by Beijing University of Aeronautics and Astronautics (application number: CN201510284550.5); the first patent discloses a psychological stress detection method based on EEG signals, including three parts: EEG signal acquisition equipment, a special EEG signal analysis system and a psychological stress level module. The working principle of this patent is: 1. The EEG signal acquisition equipment collects the wearer's brain wave signals and transmits them to a special EEG signal analysis system in real time; 2. The EEG signal is decoded and predicted 1. Process, calculate and extract the power spectral density features of EEG signals of different psychological stress levels in the database, and establish a psychological stress EEG database; 2. Apply the random forest algorithm to establish a machine learning model, classify and train the psychological stress of the characteristic values ​​of different EEG signals; 3. According to the classification model, judge the psychological stress level of the EEG signal transmitted in real time by the EEG signal acquisition device. This patent is designed based on the data features of the EEG signal, and establishes a data model of EEG signal features corresponding to different psychological stress levels, which can judge the psychological stress of the wearer of the EEG signal acquisition device in real time; the second patent is similar, and also extracts all EEG channel data and manually extracts features. First, the EEG signal is transformed by wavelet to obtain a time-frequency image, and the time-frequency graph is divided into five frequency sub-graphs of δ, Θ, α, β, and γ according to the frequency from low to high;

[0004] Next, the Gaussian mixture model (GMM) was used to estimate the probability distribution of the energy density of the time-frequency graph, and the parameters (mean, variance, weight) corresponding to the Gaussian mixture model were used as the characteristics of the EEG signal;

[0005] Then, the feature weighting algorithm (ReliefF) and support vector machine recursive feature elimination (SVM-RFE) are applied to select the above features to obtain the features that characterize the difference between normal EEG signals and epileptic EEG signals to the greatest extent. In addition, the above patents all use machine learning for feature classification.

[0006] However, these methods are inefficient in processing complex, high-dimensional EEG data and have difficulty capturing deep patterns in the data. Due to the lack of automated feature extraction capabilities, the accuracy and generalization of the models are limited, especially when dealing with individual differences and complex tasks. In addition, they all use all channels for feature learning, fail to consider individual differences, and cannot achieve personalized diagnosis. This not only increases the computational burden brought by redundant channels, but also reduces the accuracy of diagnosis.

[0007] Therefore, in view of this, the existing structure and deficiencies were studied and improved, and a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network was proposed. Summary of the invention

[0008] In view of the shortcomings of the prior art, the present invention provides a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network, which solves the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network, comprising the following steps:

[0010] S1, data preprocessing module:

[0011] Use short-time Fourier transform, hereinafter referred to as STFT, to convert the signal into a time-frequency diagram. Through STFT, the EEG signal is divided into multiple small time windows, and Fourier transform is performed in each window, thereby converting the original time domain signal into a frequency distribution at different time points. For discrete digital signals, its STFT is defined as:

[0012]

[0013] Among them, the window size is used The Hanning window is defined as follows:

[0014]

[0015] S2, adaptive channel selection module:

[0016] First, all EEG data after STFT and channel feature extraction are subjected to global average pooling and global maximum pooling operations to integrate the information of the channel feature map and generate two different one-dimensional vectors: and , respectively represent the average pooling features of all channel data and the maximum pooling features of all channel data. For the input data , indicating that The EEG data consists of channels, and its global average pooling and global max pooling It can be expressed as:

[0017]

[0018]

[0019] The two features are then forwarded to the channel weight generation module. In order to reduce parameters, fast one-dimensional convolution is used to capture cross-channel interaction information instead of using a fully connected layer, which is defined as follows:

[0020]

[0021]

[0022] The output feature vectors are then combined using element-wise summation and , and calculate the weights. In short, the channel attention feature map The calculation is as follows:

[0023]

[0024] Finally, according to the calculated attention weight, select the previous The channel features with the highest weights are weighted pooled to generate the final feature , the overall process can be described as:

[0025]

[0026]

[0027] S3, feature classification module and classification module:

[0028] The feature classification module extracts and abstractly represents the input data through two convolutional layers, and then standardizes the features through a batch normalization layer to speed up the training process and improve the generalization ability of the model. After that, the features are downsampled through a maximum pooling layer to reduce the number of parameters and the amount of calculation. Finally, the pooled features are mapped to a predefined category space through a fully connected layer to make classification decisions.

[0029] Furthermore, in step S1, short-time Fourier transform is an enhanced mathematical method derived from discrete Fourier transform, which is used to explore the instantaneous frequency and instantaneous amplitude of local waves with time-varying characteristics.

[0030] Furthermore, in step S1, discrete Fourier transform is referred to as DFT, and electroencephalogram is referred to as EEG.

[0031] Furthermore, in step S1, in formula 1, is the time domain signal to be transmitted, It's time and frequency The STFT at is a plural unit, is the window function.

[0032] Furthermore, in step S1, in formula 2, is the window function at discrete time points The value of is the length of the window.

[0033] Furthermore, in step S2, in equation 5 and equation 6, and Represents the output features.

[0034] Furthermore, in step S2, in equation 5 and equation 6, Represents a fast one-dimensional convolution with a convolution kernel of 3.

[0035] Further, in step S2, in formula 7, express Activation function, is the input data.

[0036] Furthermore, in step S2, in equation 8 and equation 9, To extract the The maximum value of the function, for Center front The maximum weight value, represents element-wise multiplication, is the original input data corresponding to the channel with the selected weight.

[0037] Furthermore, in step S2, the output of the channel selection algorithm is a set of The characteristic data of channels, The larger the value, the more channels are selected.

[0038] The present invention provides a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network, which has the following beneficial effects:

[0039] 1. This method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network converts EEG signals into time-frequency diagrams through short-time Fourier transform (STFT), effectively captures the frequency characteristics that change over time, enriches signal representation, and improves the accuracy of stress injury diagnosis. By introducing the attention mechanism, the key EEG channels related to stress injury are automatically identified and focused, achieving personalized channel selection for different individuals and improving diagnostic accuracy.

[0040] 2. This neurological stress injury diagnosis method based on EEG data channel selection and deep learning neural network combines convolutional neural network (CNN) and attention mechanism for EEG signal classification, automatically extracts complex time-frequency features, and dynamically focuses on important signal areas, overcoming the limitations of traditional manual feature extraction and fixed classification methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of the overall architecture of a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network of the present invention;

[0042] Figure 2 This is a schematic diagram of an adaptive channel selection module of a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network in the present invention. DETAILED DESCRIPTION

[0043] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0044] like Figure 1-Figure 2 As shown, the present invention provides a technical solution: a method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network, comprising the following steps:

[0045] S1, data preprocessing module:

[0046] Use short-time Fourier transform (STFT) to convert the signal into a time-frequency diagram. Through STFT, the electroencephalogram (EEG) signal is divided into multiple small time windows, and Fourier transform is performed in each window, so as to convert the original time domain signal into a frequency distribution at different time points. This time-frequency diagram can show the time and frequency changes of the signal, providing more dimensional information for EEG signal analysis. Short-time Fourier transform is an enhanced mathematical method derived from discrete Fourier transform (DFT) to explore the instantaneous frequency and instantaneous amplitude of local waves with time-varying characteristics. For discrete digital signals, its short-time Fourier transform (STFT) is defined as:

[0047]

[0048] in is the time domain signal to be transmitted, It's time and frequency The STFT at is a plural unit, is the window function. In this method, the window size is used The Hanning window is defined as follows:

[0049]

[0050] in is the window function at discrete time points The value of is the length of the window;

[0051] S2, adaptive channel selection module:

[0052] In visual tasks, the attention mechanism can be viewed as a dynamic selection process, which is achieved by adaptively weighting features according to the importance of the input. In deep learning, the attention mechanism is the ability of a model to learn how to allocate resources or focus on different parts of the model input. Since not every epilepsy data channel has the same characteristics for the recognition results, an attention mechanism is added to focus on features with more information to improve the efficiency of the system when computing resources are limited.

[0053] The existing adaptive channel selection module first extracts the features of all channels through global pooling and uses them as the input of the weight generator. The use of these two features greatly improves the representation ability of the network. The weight generator generally uses a multi-layer perceptron (MLP), which generates channel weights by using two fully connected (FC) layers and a nonlinear Sigmoid function. The dimensionality reduction operation in the MLP is likely to have a negative impact on channel attention prediction. In the fully connected layer, the model captures the correlation between all EEG channels, but this often fails to preserve the spatial structure information of the channel well, but instead increases the complexity of the model;

[0054] To solve the above problems, Figure 2 The detailed operations are as follows:

[0055] First, all EEG data after STFT and channel feature extraction are subjected to global average pooling and global maximum pooling operations to integrate the information of the channel feature map and generate two different one-dimensional vectors: and , respectively represent the average pooling features of all channel data and the maximum pooling features of all channel data. For the input data , indicating that The EEG data consists of channels, and its global average pooling and global max pooling It can be expressed as:

[0056]

[0057]

[0058] The two features are then forwarded to the channel weight generation module. In order to reduce parameters, fast one-dimensional convolution is used to capture cross-channel interaction information instead of using a fully connected layer, which is defined as follows:

[0059]

[0060]

[0061] in and represents the output features, represents a fast one-dimensional convolution with a kernel of 3, followed by the element-wise summation to combine the output feature vectors and , and calculate the weights. In short, the channel attention feature map The calculation is as follows:

[0062]

[0063] in express Activation function, is the input data,

[0064] Finally, according to the calculated attention weight, select the previous The channel features with the highest weights are weighted pooled to generate the final feature , the overall process can be described as:

[0065]

[0066]

[0067] in To extract the The maximum value of the function, for Center front The maximum weight value, represents element-wise multiplication, is the original input data corresponding to the channel with the selected weight,

[0068] The output of the channel selection algorithm is a set of The feature data of channels are considered to have high importance or relevance in predicting epileptic seizures, among which The larger the value, the more channels are selected. However, in fact, only a few key channels are needed to achieve high-accuracy diagnosis. After testing, only 30% of the original number of channels is needed to achieve the best balance between performance and computing power.

[0069] S3, feature classification module and classification module:

[0070] The feature classification module extracts and abstractly represents the input data through two convolutional layers, and then standardizes the features through a batch normalization layer to accelerate the training process and improve the generalization ability of the model. After that, the features are downsampled through a maximum pooling layer to reduce the number of parameters and the amount of calculation. Finally, the pooled features are mapped to a predefined category space through a fully connected layer to make classification decisions. No human intervention is required throughout the process, breaking away from the limitations of traditional manual feature selection. The powerful learning ability of CNN enables the model to effectively capture deep neural activity patterns and improve the efficiency and accuracy of diagnosis of stress injuries.

[0071] Existing methods usually use all EEG channels for analysis, but not every channel is directly related to the target neural activity, which causes the model to process a large number of irrelevant or redundant channel signals, thereby increasing computational complexity and resource consumption. Second, EEG signals are extremely susceptible to external interference, such as muscle movement, eye movement, environmental noise, etc. Existing methods fail to effectively remove these interfered channels, resulting in the introduction of noise in model training, affecting the accuracy of diagnostic results. In addition, high-dimensional EEG data requires a large amount of computing resources for model training and reasoning. Especially when processing deep learning models, this high computing demand is difficult to achieve on devices with limited resources, reducing real-time and practicality.

[0072] Based on this, the present invention is based on the above step description, by effectively screening out key channels that are highly correlated with target neural activity, removing irrelevant or redundant channel signals, and reducing the useless data that the model needs to process, thereby reducing computational complexity and improving efficiency. At the same time, channel reduction technology reduces the computing resource requirements for model training and reasoning by reducing the dimension of input data, which is particularly suitable for embedded devices or resource-constrained environments, ensuring the lightweight and real-time performance of the system.

[0073] Traditional methods usually use the same EEG channel configuration for all patients, failing to consider individual differences between patients, making it difficult for the model to achieve personalized and accurate channel selection, and may ignore specific channel information that is important to some patients. More importantly, existing EEG data processing methods mainly rely on manual feature extraction and traditional machine learning classifiers, such as support vector machines (SVMs) and random forests. These methods are inefficient in processing complex, high-dimensional EEG data and have difficulty capturing deep patterns in the data. Due to the lack of automated feature extraction capabilities, the accuracy and generalization ability of the model are limited, especially when dealing with individual differences and complex tasks.

[0074] Based on this, the present invention is based on the above step description, and by introducing the attention mechanism combined with the convolutional neural network (CNN) to build a classifier, it can effectively solve the problems of the prior art. The attention mechanism allows the model to dynamically focus on key features, enhance sensitivity to important information, and effectively suppress the influence of noise, thereby improving classification accuracy. Through deep learning models, especially CNN, complex EEG features are automatically extracted, which improves the efficiency and accuracy of processing high-dimensional data. In addition, this method can adapt to individual differences, realize personalized diagnosis, and improve the generalization ability of the model.

[0075] In summary, when using the neural stress injury diagnosis method based on EEG data channel selection and deep learning neural network, the overall architecture of the method is composed of four consecutive basic components, such as Figure 1As shown in the figure, these components include a data processing module, a channel feature extraction module, an adaptive channel selection module, and a classification module. The data processing module is divided into two processes. The first process is to segment the original EEG data through a sliding window of a fixed size. The second process is to apply STFT to each window to convert the EEG signal into an image spectrum. The channel feature extraction module consists of a deep separable convolution layer. The use of deep separable convolution to extract EEG data features can effectively capture spatial information, reduce the number of parameters, reduce the risk of overfitting, and retain the correlation between channels, which helps to better understand the characteristics and patterns of EEG data. The adaptive channel selection module aims to provide patients with personalized solutions. It uses the attention mechanism to assign attention weights to each channel according to the local correlation between channels. These weight scores reflect the importance of the channel, so it can ensure that the most representative channel data is effectively screened out. When selecting a channel, the channel with the highest weight score is selected. channels to ensure that the most important data is retained for subsequent analysis. This approach not only eliminates the need for prior expertise, but also significantly reduces the computational burden of subsequent data processing. The model will select channel data suitable for different patients for analysis based on their conditions, achieving adaptive channel selection. The feature classification module includes convolutional layers, batch normalization layers, pooling layers, and fully connected layers. First, the input data is feature extracted and abstracted through two convolutional layers, and then the features are standardized through a batch normalization layer to accelerate the training process and improve the generalization ability of the model. After that, the features are downsampled through a maximum pooling layer to reduce the number of parameters and the amount of computation. Finally, the pooled features are mapped to a predefined category space through a fully connected layer to make classification decisions.

[0076] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network, characterized in that: The steps include: S1, data preprocessing module: Use short-time Fourier transform, hereinafter referred to as STFT, to convert the signal into a time-frequency diagram. Through STFT, the EEG signal is divided into multiple small time windows, and Fourier transform is performed in each window, thereby converting the original time domain signal into a frequency distribution at different time points. For discrete digital signals, its STFT is defined as: ； Among them, the window size is used The Hanning window is defined as follows: ； S2, adaptive channel selection module: First, all EEG data after STFT and channel feature extraction are subjected to global average pooling and global maximum pooling operations to integrate the information of the channel feature map and generate two different one-dimensional vectors: and , respectively represent the average pooling features of all channel data and the maximum pooling features of all channel data. For the input data , indicating that The EEG data consists of channels, and its global average pooling and global max pooling It can be expressed as: ； ； The two features are then forwarded to the channel weight generation module. In order to reduce parameters, fast one-dimensional convolution is used to capture cross-channel interaction information instead of using a fully connected layer, which is defined as follows: ； ； The output feature vectors are then combined using element-wise summation and , and calculate the weights. In short, the channel attention feature map The calculation is as follows: ； Finally, according to the calculated attention weight, select the previous The channel features with the highest weights are weighted pooled to generate the final feature , the overall process can be described as: ； ； S3, feature classification module and classification module: The feature classification module extracts and abstractly represents the input data through two convolutional layers, and then standardizes the features through a batch normalization layer to speed up the training process and improve the generalization ability of the model. After that, the features are downsampled through a maximum pooling layer to reduce the number of parameters and the amount of calculation. Finally, the pooled features are mapped to a predefined category space through a fully connected layer to make classification decisions.

2. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In the step S1, short-time Fourier transform is an enhanced mathematical method derived from discrete Fourier transform, which is used to explore the instantaneous frequency and instantaneous amplitude of local waves with time-varying characteristics.

3. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In the step S1, discrete Fourier transform is referred to as DFT, and electroencephalogram is referred to as EEG.

4. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In the step S1, in formula 1, is the time domain signal to be transmitted, It's time and frequency The STFT at is a plural unit, is the window function.

5. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In step S1, in formula 2, is the window function at discrete time points The value of is the length of the window.

6. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In step S2, in equation 5 and equation 6, and Represents the output features.

7. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In step S2, in equation 5 and equation 6, Represents a fast one-dimensional convolution with a convolution kernel of 3.

8. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In step S2, in formula 7, express Activation function, is the input data.

9. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In step S2, in equation 8 and equation 9, To extract the The maximum value of the function, for Center front The maximum weight value, represents element-wise multiplication, is the original input data corresponding to the channel with the selected weight.

10. The method for diagnosing neural stress injury based on EEG data channel selection and deep learning neural network according to claim 1, characterized in that: In step S2, the output of the channel selection algorithm is a set of The characteristic data of channels, The larger the value, the more channels are selected.

Images