Two-stage electroencephalogram abnormal signal detection method
Through the two-stage EEG signal detection method, the abnormal characteristics in EEG signals are extracted and classified using autoencoder and one-dimensional convolutional neural network, the problem of difficult to identify EEG signal abnormalities in the prior art is solved, and more efficient abnormal detection and classification is achieved.
Patent Information
- Application Number
- CN202411705182.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-26
- Publication Date
- 2025-03-28
AI Technical Summary
Existing EEG signal detection algorithms are difficult to accurately identify and judge abnormal situations in diverse EEG signals, and due to the difficulty of labeling EEG data, complex signals, and long time series, normal EEG signal data and its labeling are lacking.
A two-stage EEG signal abnormality detection method is proposed. First, the high-dimensional features are extracted and compressed into low-dimensional space using an autoencoder, and the abnormal areas are detected by reconstructing the signal; then, the EEG signal time series of the candidate areas is input into the one-dimensional convolution neural network, and the features can be extracted and refined through one-dimensional convolution and depth separation convolution, and the fine classification of abnormal types is performed.
This method can accurately screen out abnormal candidate areas in the EEG signal and classify them in fine classification, which improves the accuracy and efficiency of EEG signal abnormality detection.
Smart Images
Figure CN119700137A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of bioinformatics and proposes a two-stage abnormal electroencephalogram signal detection method. Background Art
[0002] In a wide range of medical research and clinical practice, electroencephalography (EEG) plays a vital role as a key neurophysiological monitoring tool. This technology uses an array of electrodes precisely placed on the surface of the scalp to non-invasively capture and record the weak electrical signals generated by the activity of brain neuron clusters, thereby providing rich information for evaluating the functional state of the brain. In particular, in the diagnosis and evaluation of neuropsychiatric diseases, EEG has shown irreplaceable advantages, especially in the fields of epilepsy, sleep disorders and cognitive disorders. Its non-invasive, high temporal resolution and versatility make it an important medical means to evaluate, monitor and study brain function. With the continuous advancement of technology and the deepening of its application, EEG will continue to play a more important role in future medical research and clinical diagnosis and treatment.
[0003] When monitoring the patient's condition, the EEG signals of patients with neurological diseases are complex and diverse, and show different characteristics at different times. Existing detection algorithms may find it difficult to accurately identify and judge abnormalities in these diverse signals. Among the many anomaly detection methods, detection technology usually attempts to find a low-dimensional embedding of the original data, in which abnormal data and normal data are expected to be separated from each other. Therefore, it is crucial to find a suitable embedding function. However, due to the difficulty of EEG data labeling, complex signals, and lengthy time series (in the project, an EEG data contains about 20 channels, each channel has about 3 million sampling points), there is currently a lack of normal EEG signal data and its labeling. Summary of the invention
[0004] In order to solve the key problem of how to accurately locate the abnormal position of EEG waves, the present invention proposes a two-stage EEG signal abnormality detection method using autoencoder-based pre-classification and convolutional neural network, the method comprising:
[0005] Obtaining raw EEG signal data to be detected;
[0006] The time series of the original EEG signal data is input into the autoencoder model, high-dimensional features are extracted and compressed into low-dimensional space, the original data x is mapped to a low-dimensional intermediate representation h, and then the signal z is reconstructed through the low-dimensional intermediate representation h; when the output value range of the reconstructed signal z is higher than the preset threshold set based on the input vector, the data point will be marked as an abnormal signal, and the abnormal area will be intercepted and included in the candidate area;
[0007] The EEG signal time series of the candidate area is input into the convolution layer of the one-dimensional convolutional neural network, and the features are extracted and refined through one-dimensional convolution and depthwise separable convolution. Then, the probability of the features of the EEG signal time series corresponding to each category is output through the fully connected layer of the one-dimensional convolutional neural network to obtain the abnormal type.
[0008] The autoencoder model includes two parts: an encoder and a decoder;
[0009] The encoder includes a one-dimensional convolution module and a BES module, wherein the one-dimensional convolution module is used to extract high-dimensional features from the original data, and the BES module includes a down-sampling one-dimensional convolution, batch normalization, an exponential linear unit activation function, and a Dropout part, which is used to gradually embed the high-dimensional features into a low-dimensional intermediate representation;
[0010] The decoder includes a DEB module and a one-dimensional convolution module. The DEB module includes deconv1d, batch normalization, exponential linear unit activation function, and Dropout part, which are used to restore the low-dimensional intermediate representation to a high dimension. The one-dimensional convolution module is used to compress the higher-dimensional feature map to the required output dimension. The decoder maps the low-dimensional intermediate representation h back to the original input space through a structure similar to that of the encoder to generate a reconstructed high-dimensional signal z.
[0011] Calculate the mean of all data points in all channels of historical EEG signal data;
[0012] The standard deviation of the data is calculated, and the data segments within two times the standard deviation are regarded as normal EEG data and extracted from the historical EEG signal data set;
[0013] The normal EEG data is divided into multiple shorter subsequences to obtain a large number of normal EEG segment sequences.
[0014] The normal EEG segment sequence is input into the autoencoder model to train the model to minimize the reconstruction error of the normal signal.
[0015] The one-dimensional convolutional neural network includes a one-dimensional convolution module, a one-dimensional depth-separable convolution module and a fully connected layer;
[0016] The neural network includes multiple branches. The neural network first preliminarily extracts the features of the time series data of the candidate area through one-dimensional convolution, and then refines the features through deep separable convolution. The data in each branch are superimposed and fused. The fused feature map is flattened into a vector through further convolution processing, and the vector is classified through a fully connected layer with a softmax activation function.
[0017] Each branch of the neural network includes three one-dimensional convolutions and three depth-separable convolutions, and each branch uses a different convolution kernel to process the time series data of the candidate area;
[0018] The present invention has the following advantages:
[0019] 1. The present invention is based on a probabilistic statistical model and uses the distribution of amplitude to screen out EEG signals in a reasonable range to obtain normal EEG signals that can be used for training.
[0020] 2. The present invention uses normal EEG signals to train the autoencoder to reconstruct normal EEG signals, thereby obtaining an embedding function that can reconstruct normal EEG signals. The trained autoencoder is then used to reconstruct the original signal to be detected, and the difference between the input signal and the reconstructed signal is compared to detect potential anomalies, so as to screen out abnormal candidate areas.
[0021] 3. The present invention extracts and refines features through one-dimensional convolution and depthwise separable convolution on different branches, and then integrates the information extracted by different convolution kernels of each branch to form a more representative global feature map, thereby reducing the number of parameters of the model and improving the model's ability to extract specific frequency features in EEG data, and can perform in-depth analysis and classification of complex signals. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 A flowchart of a two-stage abnormal EEG signal detection method according to an embodiment of the present invention;
[0023] Figure 2 This is a structural diagram of the encoder part of the autoencoder for the two-stage abnormal EEG signal detection method according to an embodiment of the present invention;
[0024] Figure 3 This is a structural diagram of a decoder portion of an autoencoder for a two-stage abnormal EEG signal detection method according to an embodiment of the present invention;
[0025] Figure 4 This is a structural diagram of a classification model for a two-stage abnormal EEG signal detection method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0026] The best mode of implementation of the present invention is described below by way of example. It should be understood that the specific embodiments herein are used to explain the present invention in detail and should not be construed as limiting the present invention. It should be noted that various changes and modifications may be made under the premise of following the principles and core scope of the present invention, and these changes should all be deemed to be within the scope of protection of the present invention. In conjunction with the accompanying drawings, the specific implementation steps of the present invention are described in detail.
[0027] like Figure 1, which is a flow chart of a two-stage abnormal EEG signal detection method in one embodiment of the present invention. The abnormal EEG signal detection method of the present invention includes the following steps.
[0028] Step 1: Obtain the original EEG signal data to be detected;
[0029] Step 2: Input the time series of the original EEG signal data into the autoencoder model, extract high-dimensional features and compress them into low-dimensional space, map the original data x to the low-dimensional intermediate representation h, and then reconstruct the signal z through the low-dimensional intermediate representation h; when the output value range of the reconstructed signal z is higher than the preset threshold set based on the input vector, the data point will be marked as an abnormal signal, and the abnormal area will be intercepted and included in the candidate area;
[0030] In a specific embodiment, step 2 includes the following steps:
[0031] Step 2.1 Based on the probability statistics method, the overall data distribution of the EEG signal is accurately calculated to extract the normal EEG segment sequence;
[0032] Due to the complexity of EEG data annotation, signals are often highly noisy and have long time series, so it is very challenging to obtain and annotate normal EEG signal data. In particular, abnormal information in EEG signals usually exhibits sudden and short-term characteristics, which further increases the difficulty of anomaly detection.
[0033] In order to solve this problem, the present invention adopts a method based on probability statistics to accurately calculate the overall data distribution of EEG signals to extract the main components of the signal, that is, normal EEG data. Since normal EEG signals have the characteristics of regularity, stability, and low amplitude, the EEG signals in a reasonable range are screened out through the distribution of amplitude.
[0034] In a specific embodiment, the data processing step includes the following steps:
[0035] Step 2.1.1: Calculate the mean of all data points in all channels in the current individual data to determine the central trend of the data and provide a benchmark for subsequent screening.
[0036] Step 2.1.2: Calculate the standard deviation of the data and filter based on the standard deviation to obtain normal EEG data.
[0037] In a specific embodiment, data segments within two times the standard deviation are regarded as normal EEG data and extracted from the original data, which can effectively filter out potential abnormal signals and ensure the accuracy of subsequent analysis.
[0038] Step 2.1.3 divides the normal EEG data into multiple shorter subsequences to obtain a large number of normal EEG segment sequences.
[0039] Step 2.2 uses the normal EEG segment sequence to train the autoencoder model to minimize the reconstruction error of the normal EEG signal;
[0040] Among them, the autoencoder essentially simulates an identity transformation, and its main function is to maintain the characteristics of normal data when reconstructing the input data, while mapping abnormal data to different value ranges.
[0041] In a specific embodiment, the autoencoder model of the present invention is as follows Figure 2 and Figure 3 As shown, the autoencoder model consists of two parts: encoder and decoder. Figure 2 For the encoder part, Figure 3 For the decoder part;
[0042] like Figure 2 As shown in the figure, the encoder part mainly includes multiple one-dimensional convolution modules and BES modules, among which the BES module integrates downsampling one-dimensional convolution, batch normalization (Batch Normalization, BN), exponential linear unit activation function (Exponential Linear Unit, ELU), Dropout and other operations. By repeatedly passing the input data through the one-dimensional convolution and BES modules, high-dimensional features are extracted and compressed into low-dimensional space, and the input vector x is mapped to the hidden representation h. It can not only extract the high-dimensional features of the input data, but also reconstruct the original data through this simplified representation during decoding, thereby maintaining information integrity while compressing, while reducing the impact of noise and redundant information.
[0043] like Figure 3 As shown in the figure, the decoder part mainly includes multiple DEB modules and convolution modules, where the DEB module integrates deconv1d, batch normalization, exponential linear unit activation function, and Dropout operation to map the intermediate hidden representation h back to the original input space and output the reconstructed output z.
[0044] The difference between the original input vector x and the reconstructed output z is defined as the reconstruction error. In this step of the present invention, the autoencoder is trained using a normal EEG data sequence to minimize this reconstruction error. The following is the transformation of the normal EEG data sequence in the autoencoder in this step:
[0045] In a specific embodiment, normal EEG signal data first passes through a one-dimensional convolution module in the encoder, and through convolution processing, the convolution kernel slides on the data to extract features, thereby extracting local features of the time series data in the EEG data, such as amplitude, mean, variance, skewness, kurtosis, etc.
[0046] Subsequently, the data flows into the BES module, where downsampling one-dimensional convolution is first performed to extract features from the input data. Since the dimension of the downsampled data is reduced, redundant information in the data is reduced while retaining more representative features. Batch normalization is then performed to standardize each batch of data to unify the scale of the data. After the exponential linear unit activation function is used for nonlinear transformation, the information is further represented in a low-dimensional space. Finally, the Dropout operation is performed to randomly discard some features to increase the robustness of the model and prevent overfitting. These operations work together to gradually embed the original data into the intermediate hidden representation h of the high-dimensional information. The data is then subjected to one-dimensional convolution. As the convolutional network deepens, the extracted features gradually change from low-level simple features to high-level abstract features.
[0047] After the data is completely embedded in the intermediate hidden representation h of high-dimensional information, the data compressed into the low-dimensional space is first input into the DEB module, and the deconv1d operation is first performed to deconvolve the data and restore it to high dimensions. Then, batch normalization, exponential linear unit activation function, and Dropout operations are performed in sequence to map the hidden representation h back to the original input space through a transformation similar to the encoder to generate the reconstructed output z.
[0048] Step 2.3: Input the original time series data into the autoencoder model for reconstruction. When the reconstructed output value range is higher than the preset threshold, the data point will be marked as an abnormal signal, and the abnormal signal will be intercepted and included in the candidate area.
[0049] In a specific embodiment, by training the autoencoder on a normal time series data set, it is possible to learn and reconstruct normal patterns in the data. Since the autoencoder does not learn the characteristics of abnormal data, the decoder is usually unable to restore the original input when reconstructing the abnormal data, and the result may deviate from the input value and map to a different value range.
[0050] After training, the autoencoder can be used to reconstruct the original time series data. When the reconstructed output value range is higher than a preset threshold, the data point will be marked as an abnormal signal. At this time, by intercepting these abnormal areas, candidate areas can be generated to prepare for the subsequent fine classification stage.
[0051] The autoencoder not only improves the ability to distinguish between normal and abnormal signals, but also provides a solid foundation for accurately locating abnormal brain wave locations.
[0052] Step 3: Input the EEG signal time series of the candidate area into the convolution layer of the one-dimensional convolutional neural network, extract and refine the features through one-dimensional convolution and depthwise separable convolution, and then output the probability of the features of the EEG signal time series corresponding to each category through the fully connected layer of the one-dimensional convolutional neural network to obtain the abnormal type.
[0053] like Figure 4 , which is a schematic diagram of a one-dimensional convolutional neural network according to an embodiment of the present invention. The classification model includes a convolutional layer, a fusion layer and a fully connected layer.
[0054] The convolution layer includes multiple branches, each of which includes three one-dimensional convolutions and three depth-separable convolutions. Each branch uses a different convolution kernel to process the data, and batch normalization (BN) and ELU activation function operations are set after each convolution. BN can speed up the training process and improve the stability of the model. The ELU activation function has more advantages than ReLU when processing negative values, which helps the model better capture the nonlinear characteristics of the data.
[0055] Among them, the one-dimensional convolution module is used to preliminarily extract the electric wave features of the EEG signal time series data, while the deep separable convolution further refines these features.
[0056] Depthwise separable convolution operates by decomposing the standard convolution into two parts: depthwise convolution and pointwise convolution. First, depthwise convolution is performed (convolution is performed on each input channel independently), and then pointwise convolution is performed (1x1 convolution is performed on the output of depthwise convolution to mix information from different channels). This method significantly reduces the number of parameters of the model, while improving the model's ability to extract the temporal correlation features between each sampling point in a single specific frequency in the EEG data, and can better identify the characteristics of specific frequencies.
[0057] In EEG signals, changes in certain frequencies may be associated with specific EEG activities or states. The design of deep separable convolution allows the model to capture these frequency features while taking into account the relationships between each time point, thereby improving the model's ability to understand these complex temporal features.
[0058] Among them, the fusion layer includes cross-scale fusion and convolution operations. The data in each independent branch are first superimposed and cross-scale fused. The fused data is processed by three one-dimensional convolutions. After each convolution, batch normalization (BN) and ELU activation function operations are performed, and the obtained feature maps are fused and flattened into a vector.
[0059] Among them, the fully connected layer includes a softmax activation function and a dropout operation. The softmax activation function is used to output the probability of each category to achieve accurate classification of abnormal types. Dropout prevents model overfitting by randomly discarding some neurons.
[0060] The present invention greatly improves the classification accuracy of abnormal areas of EEG signals through optimized neural network structure design.
[0061] In a specific embodiment, after completing the preliminary abnormality detection in the first stage, all sequences screened as candidate areas are input into a fine classification model for further processing. The abnormal EEG data after pre-classification is first subjected to three one-dimensional convolutions and three depth-separable convolutions with different convolution kernels to extract abnormal features, and then the data in each independent branch is superimposed and cross-scale fused to obtain multi-scale abnormal features, integrating the information of each branch to form a more representative global feature map. After the initial fusion, further convolution operations are performed to extract deeper features based on the fused features, and further compress and refine the information. Finally, the obtained feature map is flattened into a vector, and then classified through a fully connected layer with a softmax activation function to output the category of the abnormal EEG signal.
Claims
1. A two-stage abnormal EEG signal detection method comprising: Obtaining raw EEG signal data to be detected; The time series of the original EEG signal data is input into the autoencoder model, high-dimensional features are extracted and compressed into low-dimensional space, the original data x is mapped to a low-dimensional intermediate representation h, and then the signal z is reconstructed through the low-dimensional intermediate representation h; when the output value range of the reconstructed signal z is higher than the preset threshold set based on the input vector, the data point will be marked as an abnormal signal, and the abnormal area will be intercepted and included in the candidate area; The EEG signal time series of the candidate area is input into a one-dimensional convolutional neural network. The features are extracted and refined through one-dimensional convolution and depthwise separable convolution. After the features are fused, the softmax activation function is used to output the probability of the features of the EEG signal time series corresponding to each category to obtain the abnormal type.
2. The method according to claim 1, characterized in that: The autoencoder model includes two parts: an encoder and a decoder; The encoder includes a one-dimensional convolution module and a BES module, wherein the one-dimensional convolution module is used to extract high-dimensional features from the original data, and the BES module includes a down-sampling one-dimensional convolution, batch normalization, an exponential linear unit activation function, and a Dropout part, which is used to gradually embed the high-dimensional features into a low-dimensional intermediate representation; The decoder includes a DEB module and a one-dimensional convolution module. The DEB module includes deconv1d, batch normalization, an exponential linear unit activation function and a Dropout part, which is used to restore the low-dimensional intermediate representation to a high dimension. The one-dimensional convolution module is used to compress the higher-dimensional feature map to the required output dimension. The decoder maps the low-dimensional intermediate representation h back to the original input space through a structure similar to that of the encoder to generate a reconstructed high-dimensional signal z.
3. The method according to claim 1, characterized in that Also includes: Calculate the mean of all data points in all channels of historical EEG signal data; The standard deviation of the data is calculated, and the data segments within two times the standard deviation are regarded as normal EEG data and extracted from the historical EEG signal data set; The normal EEG data is divided into multiple shorter subsequences to obtain a large number of normal EEG segment sequences.
4. The method according to claim 1, characterized in that Also includes: The normal EEG segment sequence is input into the autoencoder model to train the model to minimize the reconstruction error of the normal signal.
5. The method according to claim 1, characterized in that: The one-dimensional convolutional neural network includes a convolutional layer, a fusion layer and a fully connected layer; The convolution layer includes multiple branches. The features of the time series data of the candidate area are first extracted through one-dimensional convolution, and then the features are refined through deep separable convolution. The fusion layer superimposes the data in each branch for cross-scale fusion processing. The fused feature map is convolved and flattened into a vector. The fully connected layer classifies the vector through a softmax activation function.
6. The method according to claim 5, characterized in that Each branch of the convolution layer includes a one-dimensional convolution and a depth-separable convolution, and each branch uses a different convolution kernel to process the time series data of the candidate area.