Electroencephalogram signal visualization channel selection method and device for autism

Abstract

The application provides an electroencephalogram (EEG) visualization channel selection method and device for autism, which comprises the following steps: preprocessing original EEG data; training, hyperparameter optimization and performance testing of a deep convolutional neural network model based on the preprocessed EEG data; obtaining a class activation map corresponding to each input sample based on a plurality of feature maps corresponding to each input sample generated by a last convolutional layer of the deep convolutional neural network model whose performance meets an expectation; determining an importance score of each channel according to the class activation map corresponding to each input sample; and sorting the channels in descending order of the importance scores and selecting channels ranked in the first preset positions. The application provides guidance for selecting EEG electrodes for measuring autism spectrum disorder subjects, effectively reduces the measurement cost, improves the test efficiency, and reduces the burden and stress of the subjects.

Description

Technical Field

[0001] The present invention relates to the field of electroencephalogram (EEG) technology, and in particular to a method, device, and readable storage medium for selecting an EEG signal visualization channel for autism. Background Art

[0002] Autism spectrum disorder is a broadly defined autism based on the core symptoms of typical autism. Its symptoms include social communication disorders, language communication disorders, and repetitive stereotyped behaviors. There are a large number of patients with autism spectrum disorder in my country, and early screening and identification are extremely important. However, the current assessment and diagnosis technology for autism in my country is mainly based on expert interviews and behavioral scales, which is still subjective, extensive, and inefficient. Electroencephalography (EEG) is a recording of electrophysiological activity in the cerebral cortex. It has the characteristics of non-invasiveness, acquisition, and high temporal resolution. Combined with feature engineering and traditional machine learning technology, it can realize the automatic identification of autistic people. Its technical path is: (1) Preprocessing. The original acquired signal is filtered to retain the signal components of a specific frequency band and remove artifacts in the signal; (2) Feature extraction. Through signal processing, nonlinear measurement and other analysis methods, time, frequency, entropy and other features are extracted from the signal to form a feature set; (3) Feature selection. Use feature selection algorithms based on filtering, wrapping, etc. to evaluate feature importance and select effective feature subsets; (4) Autism discrimination. Select a classifier or set of classifiers and optimize the classifier parameters using a set of labeled samples. Use the optimized model to discriminate new unlabeled samples. The advantage of this solution is that the training complexity is low and it is relatively independent of data and computing power; the disadvantage is that it requires a high level of prior knowledge and professional skills.

[0003] With the continuous accumulation of existing big data and the continuous improvement of computing power, deep learning technologies, represented by convolutional neural networks and recurrent neural networks, have recently begun to be widely applied in fields such as computer vision and natural language processing, and have even achieved remarkable results far exceeding human performance under certain conditions. Therefore, combining deep learning technologies with EEG signals to achieve high-performance and accurate autism detection has become a hot topic in the field of brain health. However, high-density EEG acquisition equipment is expensive and places a greater burden on subjects (especially children with autism), significantly increasing the cost and difficulty of EEG signal acquisition. Signals from some channels are not helpful for autism diagnosis and can actually reduce the efficiency and performance of model training. Furthermore, end-to-end deep neural network models often have complex and nonlinear internal structures, resulting in a lack of interpretability of detection results, which severely limits the practical application of this technology. Summary of the Invention

[0004] To solve at least one of the above technical problems, the present invention provides a method, device and readable storage medium for selecting EEG signal visualization channels for autism.

[0005] In a first aspect, the present invention provides a method for selecting a channel for visualizing an EEG signal for autism, characterized in that the method comprises:

[0006] Obtaining raw EEG data collected by an EEG data acquisition device;

[0007] Preprocess the raw EEG data;

[0008] Based on the preprocessed EEG data, the deep convolutional neural network model is trained, hyperparameters are optimized, and performance is tested until a deep convolutional neural network model with expected performance is obtained;

[0009] Obtaining a class activation map corresponding to each input sample based on multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance;

[0010] Determine the importance score of each channel based on the class activation map corresponding to each input sample;

[0011] Sort the channels in descending order of importance score, and select the channels that are pre-positioned at the top.

[0012] Optionally, the step of preprocessing the raw EEG data includes:

[0013] The raw EEG data were denoised using downsampling, bandpass filtering and artifact removal algorithms;

[0014] Use a fixed time window to split the denoised raw EEG data into multiple EEG data segments;

[0015] Performing discrete Fourier transform on each EEG data segment to obtain multiple input samples;

[0016] The multiple input samples are divided into a training set, a validation set and a test set according to a preset ratio, wherein the input samples belonging to the same individual are in the same data set.

[0017] Optionally, the input sample is three-dimensional tensor EEG data; the hidden layer of the deep convolutional neural network model is composed of a convolutional layer, a maximum pooling layer, an average pooling layer and a dropout layer, wherein the convolutional layer, the maximum pooling layer and the average pooling layer all adopt a two-dimensional structure; the output layer of the deep convolutional neural network model is a fully connected layer with 2 nodes, and is connected to a soft maximum Softmax activation function, so that the output layer outputs the probability that the sample belongs to autism and the probability of a normal individual, which is 100%.

[0018] Optionally, the step of obtaining a class activation map corresponding to each input sample based on the multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model that achieves expected performance includes:

[0019] For a given input sample, obtain the feature map [n × a × b] output by the last convolutional layer of the deep convolutional neural network model that achieves the expected performance, where n is the number of feature maps and a × b is the size of each feature map;

[0020] calculate Return the gradient information of the feature layer :

[0021]

[0022] in, represents the prediction score of the network layer before the softmax layer for category c, is the value of the kth channel of the nth feature map A at position (i, j);

[0023] Calculate the importance weight of the feature map for category c , where Z represents the scale of the feature layer, which is the product of the number of channels and the bandwidth:

[0024]

[0025] Each pair of weights is weighted and linearly combined to form a feature map, and then fed into the ReLU activation function to obtain the class activation map:

[0026] .

[0027] Optionally, after the step of weighting each pair of weight pairs and linearly combining the feature maps and inputting the weights into a ReLU activation function to obtain a class activation map, the method further includes:

[0028] The class activation map is normalized and upsampled to [a × b] size, and then superimposed on the corresponding input sample using a color map for display. The weight of each channel in each frequency band is recorded.

[0029] Optionally, the step of determining the importance score of each channel according to the class activation map corresponding to each input sample includes:

[0030] Process the class activation map of each input sample, accumulate the weights between different frequency bands of each channel, and obtain the total weight of the i-th channel of the n-th image ;

[0031] For the i-th channel of the n-th image, if the total weight is higher than the importance threshold T, the importance score of this channel is increased, otherwise no processing is done, and the importance score of each channel is obtained. :

[0032] ;

[0033] in, .

[0034] Optionally, after the step of sorting the channels in descending order of importance scores and selecting the channels ranked in the top preset positions, the method further includes:

[0035] Obtaining target EEG data corresponding to the channel at the front preset position from the pre-processed EEG data;

[0036] The deep convolutional neural network model is trained and hyperparameters are optimized based on the target EEG data to obtain a new deep convolutional neural network model;

[0037] Performing performance testing on the new deep convolutional neural network model;

[0038] The channel selection accuracy is determined based on the performance test results.

[0039] In a second aspect, the present invention further provides a device for selecting a channel for visualizing an EEG signal for autism, the device comprising:

[0040] An acquisition module is used to acquire the original EEG data collected by the EEG data acquisition device;

[0041] Preprocessing module, used to preprocess the raw EEG data;

[0042] The training module is used to train the deep convolutional neural network model based on the preprocessed EEG data, optimize hyperparameters, and perform performance testing until a deep convolutional neural network model with expected performance is obtained;

[0043] A generation module is configured to obtain a class activation map corresponding to each input sample based on multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance;

[0044] The scoring module is used to determine the importance score of each channel based on the class activation map corresponding to each input sample;

[0045] The screening module is used to sort the channels in descending order of importance scores and select the channels with the highest preset positions.

[0046] In a third aspect, the present invention also provides an EEG signal visualization channel selection device for autism, wherein the EEG signal visualization channel selection device for autism comprises a processor, a memory, and an EEG signal visualization channel selection program for autism stored on the memory and executable by the processor, wherein when the EEG signal visualization channel selection program for autism is executed by the processor, the steps of the EEG signal visualization channel selection method for autism as described above are implemented.

[0047] In a fourth aspect, the present invention also provides a readable storage medium, on which is stored a program for selecting a channel for visualizing EEG signals for autism. When the program for visualizing EEG signals for autism is executed by a processor, the steps of the method for selecting a channel for visualizing EEG signals for autism as described above are implemented.

[0048] In the present invention, the raw EEG data collected by the EEG data acquisition device is obtained; the raw EEG data is preprocessed; based on the preprocessed EEG data, a deep convolutional neural network model is trained, hyperparameters are optimized, and performance tested until a deep convolutional neural network model with expected performance is obtained; based on the multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance, a class activation map corresponding to each input sample is obtained; based on the class activation map corresponding to each input sample, the importance score of each channel is determined; the channels are sorted in descending order according to the importance score, and the channels with the highest preset positions are selected. The present invention has the following positive effects and advantages:

[0049] 1. The present invention can obtain a class activation map from the existing deep neural network discriminant model by combining the final feature map with the score of the target category without the need for additional structural changes or retraining, thereby generating a feature contribution map together with the sample, and the process does not rely on computing resources or a posteriori knowledge.

[0050] 2. This invention addresses, to some extent, the lack of interpretability and discriminative evidence when using end-to-end deep learning models. Furthermore, by comparing feature contribution graphs across individuals, we can further explore the individual differences among patients with autism spectrum disorders.

[0051] 3. This invention provides guidance for subsequent selection of EEG electrodes for measurement in subjects with autism spectrum disorders through channel selection, effectively reducing measurement costs, improving testing efficiency, and alleviating the burden and stress on subjects. Furthermore, this invention can effectively improve model calculation speed and performance by rationally and accurately reducing the number of channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of an embodiment of a method for selecting channels for visualizing EEG signals for autism according to the present invention;

[0053] Figure 2 Schematic diagram of a channel screening scenario in one embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the hardware structure of the EEG signal visualization channel selection device for autism involved in the embodiment of the present invention.

[0055] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0056] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0057] In a first aspect, an embodiment of the present invention provides a method for selecting EEG signal visualization channels for autism.

[0058] In one embodiment, referring to Figure 1 , Figure 1 FIG. 1 is a flow chart of an embodiment of a method for selecting a channel for visualizing an EEG signal for autism according to the present invention. Figure 1 As shown in FIG, the EEG signal visualization channel selection method for autism includes:

[0059] Step S10, obtaining the original EEG data collected by the EEG data acquisition device;

[0060] In this embodiment, the raw EEG data of the subject is collected using an EEG data acquisition device according to a standard acquisition process. Specifically, EEG signal acquisition is usually carried out in a professional EEG acquisition room. Sensors are attached to the surface of the human scalp to record the data of the person at rest or in a task state, and the data is transmitted to a computer in real time for storage. The subject is required to remove glasses, earrings, and other objects and wear a headgear made of special materials. Electrodes corresponding to 62 channels, a reference electrode, and a ground wire are passed through the headgear to fit tightly to the subject's scalp, and a special conductive paste is applied to the contact position to reduce the resistivity. The subject is asked to maintain a still, open-eyed state with steady breathing for 3-5 minutes. The EEG signals generated during this period are collected and stored together with the subject's information.

[0061] Step S20, preprocessing the original EEG data;

[0062] In this embodiment, the preprocessing mainly includes denoising, segmentation and discrete Fourier transform processing.

[0063] Furthermore, in one embodiment, step S20 includes:

[0064] The raw EEG data is denoised using downsampling, bandpass filtering and artifact removal algorithms; the denoised raw EEG data is segmented into multiple EEG data segments using a fixed time window; each EEG data segment is discrete Fourier transformed to obtain multiple input samples; the multiple input samples are divided into a training set, a validation set and a test set according to a preset ratio, where input samples belonging to the same individual are in the same data set.

[0065] In this embodiment, the default frequency of the original EEG data is 1000Hz, which is much higher than the resolution requirement of the general deep neural network for data. At the same time, in order to improve its noise resistance, the data is reduced to 256Hz by using the downsampling method. Next, a bandpass filtering algorithm is used to filter out the parts between 0-0.5Hz and above 45Hz. Each denoised original EEG data is divided into 4-second EEG data segments as the unit for subsequent processing. For segments with a tail of less than 4s when the signal is segmented, the segment is discarded. In the subsequent discrimination, the discrimination results of more than half of all the segments of the individual are used as the discrimination results of the individual. Next, the segments with more noise or drift are manually identified and discarded. Each EEG data segment is subjected to discrete Fourier transform processing, and it is converted from the time domain to the frequency domain to make the features more prominent, while only the 0.5-45Hz portion is retained again. To facilitate subsequent processing, each sample was upscaled from two dimensions to three dimensions, resulting in a final size of [1 × 62 × 182], where 62 corresponds to 62 channels and 182 corresponds to the length of the 0.5-45 Hz segment. Finally, all samples, along with their corresponding labels (each sample's label format was [0, 1] or [1, 0], corresponding to individuals with autism and typically developing individuals, respectively), were randomly mixed by individual (rather than by segment) and divided into training, validation, and test sets in a 3:1:1 ratio. The first two were used for model training and optimization, while the latter was used for model testing.

[0066] Step S30, training, hyperparameter optimization, and performance testing the deep convolutional neural network model based on the preprocessed EEG data until a deep convolutional neural network model with expected performance is obtained;

[0067] In this example, the model (i.e., a deep convolutional neural network model) takes as input samples [n × 1 × 62 × 182] and labels [n × 2]. The output is an array of length 2, where each value is between 0 and 1 and sums to 1, representing the probability of autism and normality, respectively. The specific model structure can be adjusted and optimized based on experience and verification accuracy, but the following principles should be followed:

[0068] (1) The main part of the model (front center) should be composed of alternating convolutional layers, maximum pooling layers, and average pooling layers, with a depth between 4 and 8;

[0069] (2) Considering the data size, the number of convolution kernels in the convolution layer should be between 8 and 64. To ensure that the last feature map still retains 62 channels, the window size of the convolution layer and the pooling layer should be between 1×3 and 1×7. The convolution layer step size is recommended to be 1×1, and the pooling layer step size is recommended to be consistent with the window size.

[0070] (3) After each convolution / pooling layer, activation functions and dropout layers can be added as appropriate. ReLU is recommended for the former, and the dropout probability of the latter is recommended to be 0.2-0.4;

[0071] (4) At the end of the front middle part, the data size is flattened from three dimensions to one dimension through the flattening layer, thus connecting to the subsequent fully connected layer;

[0072] (5) The depth of the fully connected layer is 2 layers. The number of cores should be from large to small. The number of cores in the last layer is 2 (corresponding to the size of the result). The activation function is softmax. The other activation functions are recommended to be ReLU.

[0073] The training parameters follow the following principles:

[0074] (1) To speed up training, it is recommended to use small batches, with the batch size between 128 and 512 depending on the size of the dataset;

[0075] (2) The Loss function should be categorical cross entropy.

[0076] (3) The optimizer recommends choosing the ADAM optimizer that can dynamically adjust the learning rate.

[0077] (4) The recommended training cycle is 100, and the system is set to automatically stop early based on the validation error and save the optimal parameters. After the model is trained and optimized, the final performance of the model can be evaluated using the test dataset, including test accuracy, model specificity, and model sensitivity.

[0078] Furthermore, in one embodiment, the input sample is three-dimensional tensor EEG data; the hidden layer of the deep convolutional neural network model is composed of a convolutional layer, a maximum pooling layer, an average pooling layer and a dropout layer, wherein the convolutional layer, the maximum pooling layer and the average pooling layer all adopt a two-dimensional structure; the output layer of the deep convolutional neural network model is a fully connected layer with 2 nodes, and is connected to a soft maximum Softmax activation function, so that the output layer outputs the probability that the sample belongs to autism and the probability of a normal individual, which is 100%.

[0079] In step S40, a class activation map corresponding to each input sample is obtained based on multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance; in step S50, an importance score of each channel is determined according to the class activation map corresponding to each input sample; in step S60, the channels are sorted in descending order according to the importance score, and the channels with the front preset position are selected.

[0080] Furthermore, in one embodiment, step S40 includes:

[0081] For a given input sample, obtain the feature map [n × a × b] output by the last convolutional layer of the deep convolutional neural network model that achieves the expected performance, where n is the number of feature maps and a × b is the size of each feature map;

[0082] calculate Return the gradient information of the feature layer :

[0083]

[0084] in, represents the prediction score of the network layer before the softmax layer for category c, is the value of the kth channel of the nth feature map A at position (i, j);

[0085] Calculate the importance weight of the feature map for category c , where Z represents the scale of the feature layer, which is the product of the number of channels and the bandwidth:

[0086]

[0087] Each pair of weights is weighted and linearly combined to form a feature map, and then fed into the ReLU activation function to obtain the class activation map:

[0088] .

[0089] Reference Figure 2 , Figure 2 FIG. 1 is a schematic diagram of a channel screening scenario in an embodiment of the present invention. Figure 2 As shown in the figure, for a certain input sample, obtain the feature map [n × a × b] output by the last convolutional layer of the deep convolutional neural network model with the expected performance, where n is the number of feature maps and a × b is the size of each feature map. Among them, a is fixed to 62, and n and b are determined by the neural network structure and hyperparameters set previously. Next, obtain , that is, the prediction score of the network layer for category c before the softmax layer. Calculate the gradient information of each feature map back to the feature layer . Then calculate the importance weight of the feature map for category c. Next, weight each pair of weights on the feature map and linearly combine them, and send them to the ReLu activation function to obtain the class activation map. The size of this class activation map is the same as that of a single feature map [a×b]. In order to superimpose with the original sample, the class activation map should be upsampled to the same size as a single sample [62× 182] (it is recommended to use interpolation for upsampling). Each point in the final class activation map is the contribution of the feature at the corresponding position in the corresponding input sample to the discrimination result c (the range is 0-1, no unit, the higher the contribution, the greater the contribution).

[0090] Furthermore, in one embodiment, step S50 includes:

[0091] Process the class activation map of each input sample, accumulate the weights between different frequency bands of each channel, and obtain the total weight of the i-th channel of the n-th image ;

[0092] For the i-th channel of the n-th image, if the total weight is higher than the importance threshold T, the importance score of this channel is increased, otherwise no processing is done, and the importance score of each channel is obtained. :

[0093] ;

[0094] in, .

[0095] In this embodiment, the weight values ​​in the class activation map of each input sample are obtained and the weight values ​​of each class activation map are accumulated according to the channel dimension to form the total weight of the channel. Next, an importance threshold T is set and an importance score is assigned to each channel. For the i-th channel in the n-th image, if the total weight is greater than T, the importance score of this channel is increased; otherwise, no processing is performed. Finally, the importance scores of all channels are sorted from highest to lowest, and the top m channels are selected based on the number of channels to be screened.

[0096] Furthermore, in one embodiment, after step S40, the following steps are further included:

[0097] The class activation map is normalized and upsampled to [a × b] size, and then superimposed on the corresponding input sample using a color map for display. The weight of each channel in each frequency band is recorded.

[0098] Furthermore, in one embodiment, after step S60, the method further includes:

[0099] Obtain target EEG data corresponding to the front preset channel from the preprocessed EEG data; train a deep convolutional neural network model and optimize hyperparameters based on the target EEG data to obtain a new deep convolutional neural network model; perform performance testing on the new deep convolutional neural network model; and determine the accuracy of channel selection based on the performance test results.

[0100] This embodiment has the following positive effects and advantages:

[0101] This embodiment can obtain a class activation map from the existing deep neural network discriminant model by combining the final feature map with the score of the target category without the need for additional structural changes or retraining, thereby generating a feature contribution map together with the sample, and the process does not need to rely on computing resources or a posteriori knowledge.

[0102] This embodiment, to some extent, addresses the lack of interpretability and discriminative evidence when using end-to-end models such as deep learning. Furthermore, individual differences in autism spectrum disorder patients can be further explored by comparing feature contribution graphs between different individuals.

[0103] This embodiment provides guidance for selecting subsequent EEG electrodes for measurement in subjects with autism spectrum disorder through channel selection, effectively reducing measurement costs, improving testing efficiency, and alleviating the burden and stress on subjects. Furthermore, this invention can effectively improve model calculation speed and performance by rationally and accurately reducing the number of channels.

[0104] In a second aspect, an embodiment of the present invention provides a device for selecting a channel for visualizing an EEG signal for autism, the device comprising:

[0105] An acquisition module is used to acquire the original EEG data collected by the EEG data acquisition device;

[0106] Preprocessing module, used to preprocess the raw EEG data;

[0107] The training module is used to train the deep convolutional neural network model based on the preprocessed EEG data, optimize hyperparameters, and perform performance testing until a deep convolutional neural network model with expected performance is obtained;

[0108] A generation module is configured to obtain a class activation map corresponding to each input sample based on multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance;

[0109] The scoring module is used to determine the importance score of each channel based on the class activation map corresponding to each input sample;

[0110] The screening module is used to sort the channels in descending order of importance scores and select the channels with the highest preset positions.

[0111] Furthermore, in one embodiment, the pre-processing module is configured to:

[0112] The raw EEG data were denoised using downsampling, bandpass filtering and artifact removal algorithms;

[0113] Use a fixed time window to split the denoised raw EEG data into multiple EEG data segments;

[0114] Performing discrete Fourier transform on each EEG data segment to obtain multiple input samples;

[0115] The multiple input samples are divided into a training set, a validation set and a test set according to a preset ratio, wherein the input samples belonging to the same individual are in the same data set.

[0116] Furthermore, in one embodiment, the input sample is three-dimensional tensor EEG data; the hidden layer of the deep convolutional neural network model is composed of a convolutional layer, a maximum pooling layer, an average pooling layer and a dropout layer, wherein the convolutional layer, the maximum pooling layer and the average pooling layer all adopt a two-dimensional structure; the output layer of the deep convolutional neural network model is a fully connected layer with 2 nodes, and is connected to a soft maximum Softmax activation function, so that the output layer outputs the probability that the sample belongs to autism and the probability of a normal individual, which is 100%.

[0117] Furthermore, in one embodiment, the generation module is configured to:

[0118] For a given input sample, obtain the feature map [n × a × b] output by the last convolutional layer of the deep convolutional neural network model that achieves the expected performance, where n is the number of feature maps and a × b is the size of each feature map;

[0119] calculate Return the gradient information of the feature layer :

[0120]

[0121] in, represents the prediction score of the network layer before the softmax layer for category c, is the value of the kth channel of the nth feature map A at position (i, j);

[0122] Calculate the importance weight of the feature map for category c , where Z represents the scale of the feature layer, which is the product of the number of channels and the bandwidth:

[0123]

[0124] Each pair of weights is weighted and linearly combined to form a feature map, and then fed into the ReLU activation function to obtain the class activation map:

[0125] .

[0126] Furthermore, in one embodiment, the EEG signal visualization channel selection device for autism further includes a display module for:

[0127] The class activation map is normalized and upsampled to [a × b] size, and then superimposed on the corresponding input sample using a color map for display. The weight of each channel in each frequency band is recorded.

[0128] Furthermore, in one embodiment, the scoring module is used to:

[0129] Process the class activation map of each input sample, accumulate the weights between different frequency bands of each channel, and obtain the total weight of the i-th channel of the n-th image ;

[0130] For the i-th channel of the n-th image, if the total weight is higher than the importance threshold T, the importance score of this channel is increased, otherwise no processing is done, and the importance score of each channel is obtained. :

[0131] ;

[0132] in, .

[0133] Furthermore, in one embodiment, the apparatus for selecting a channel for visualizing an EEG signal for autism further includes a verification module for:

[0134] Obtaining target EEG data corresponding to the channel at the front preset position from the pre-processed EEG data;

[0135] The deep convolutional neural network model is trained and hyperparameters are optimized based on the target EEG data to obtain a new deep convolutional neural network model;

[0136] Performing performance testing on the new deep convolutional neural network model;

[0137] The channel selection accuracy is determined based on the performance test results.

[0138] The specific embodiments of the device for selecting a visual channel for EEG signals for autism according to the present invention refer to the various embodiments of the method for selecting a visual channel for EEG signals for autism described above, and are not described in detail here.

[0139] In a third aspect, an embodiment of the present invention provides an EEG signal visualization channel selection device for autism. The EEG signal visualization channel selection device for autism can be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.

[0140] Reference Figure 3 , Figure 3 This is a hardware structure diagram of the EEG signal visualization channel selection device for autism involved in the embodiment of the present invention. In the embodiment of the present invention, the EEG signal visualization channel selection device for autism may include a processor 1001 (such as a central processing unit, CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components; the user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard); the network interface 1004 may optionally include a standard wired interface, a wireless interface (such as a wireless fidelity WIreless-FIdelity, WI-FI interface); the memory 1005 may be a high-speed random access memory (RAM), or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001. Those skilled in the art will understand that Figure 3 The hardware structure shown in the figure does not constitute a limitation of the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0141] Continue to refer to Figure 3 , Figure 3 The memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a program for selecting a channel for visualizing EEG signals for autism. The processor 1001 may call the program for selecting a channel for visualizing EEG signals for autism stored in the memory 1005 and execute the method for selecting a channel for visualizing EEG signals for autism provided in an embodiment of the present invention.

[0142] In a fourth aspect, an embodiment of the present invention further provides a readable storage medium.

[0143] The readable storage medium of the present invention stores a program for selecting a channel for visualizing EEG signals for autism, wherein when the program for visualizing EEG signals for autism is executed by a processor, the steps of the method for visualizing EEG signals for autism are implemented.

[0144] Among them, the method implemented when the EEG signal visualization channel selection program for autism is executed can refer to the various embodiments of the EEG signal visualization channel selection method for autism of the present invention, and will not be repeated here.

[0145] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0146] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0147] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, or of course by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device to execute the methods described in various embodiments of the present invention.

[0148] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for selecting channels for visualizing EEG signals for autism, characterized in that: The method for selecting a channel for visualizing an EEG signal for autism includes: Obtaining raw EEG data collected by an EEG data acquisition device; Preprocess the raw EEG data; Based on the preprocessed EEG data, the deep convolutional neural network model is trained, hyperparameters are optimized, and performance is tested until a deep convolutional neural network model with expected performance is obtained; Obtaining a class activation map corresponding to each input sample based on multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance; Determine the importance score of each channel based on the class activation map corresponding to each input sample; Sort the channels in descending order of importance and select the channels that are ranked first; The step of obtaining a class activation map corresponding to each input sample based on multiple feature maps corresponding to each input sample generated by the last convolutional layer of the deep convolutional neural network model with expected performance includes: For a given input sample, obtain the feature map [n × a × b] output by the last convolutional layer of the deep convolutional neural network model that achieves the expected performance, where n is the number of feature maps and a × b is the size of each feature map; calculate Return the gradient information of the feature layer : in, represents the prediction score of the network layer before the softmax layer for category c, is the value of the kth channel of the nth feature map A at position (i, j); Calculate the importance weight of the feature map for category c , where Z represents the scale of the feature layer, which is the product of the number of channels and the bandwidth: Each pair of weights is weighted and linearly combined to form a feature map, and then fed into the ReLU activation function to obtain the class activation map: ； The step of determining the importance score of each channel according to the class activation map corresponding to each input sample includes: Process the class activation map of each input sample, accumulate the weights between different frequency bands of each channel, and obtain the total weight of the i-th channel of the n-th image ; For the i-th channel of the n-th image, if the total weight is higher than the importance threshold T, the importance score of this channel is increased, otherwise no processing is done, and the importance score of each channel is obtained. : ； in, .

2. The method for selecting a channel for visualizing an EEG signal for autism according to claim 1, wherein: The step of preprocessing the raw EEG data comprises: The raw EEG data were denoised using downsampling, bandpass filtering and artifact removal algorithms; Use a fixed time window to split the denoised raw EEG data into multiple EEG data segments; Performing discrete Fourier transform on each EEG data segment to obtain multiple input samples; The multiple input samples are divided into a training set, a validation set and a test set according to a preset ratio, wherein the input samples belonging to the same individual are in the same data set.

3. The method for selecting a channel for visualizing an EEG signal for autism according to claim 2, wherein: The input sample is three-dimensional tensor EEG data; the hidden layer of the deep convolutional neural network model includes a convolution layer, a maximum pooling layer, an average pooling layer and a dropout layer, wherein the convolution layer, the maximum pooling layer and the average pooling layer all adopt a two-dimensional structure; the output layer of the deep convolutional neural network model is a fully connected layer with 2 nodes, and is connected to a soft maximum Softmax activation function, so that the output layer outputs the probability that the sample belongs to autism and the probability of a normal individual, which is 100%.

4. The method for selecting a channel for visualizing an EEG signal for autism according to claim 1, wherein: After the step of weighting each pair of weights and linearly combining the feature maps and inputting the weights into a ReLU activation function to obtain a class activation map, the method further includes: The class activation map is normalized and upsampled to [a × b] size, and then superimposed on the corresponding input sample using a color map for display. The weight of each channel in each frequency band is recorded.

5. The method for selecting a channel for visualizing an EEG signal for autism according to claim 1, wherein: After the step of sorting the channels in descending order of importance scores and selecting the channels ranked at the top, the method further includes: Obtaining target EEG data corresponding to the channel at the front preset position from the pre-processed EEG data; The deep convolutional neural network model is trained and hyperparameters are optimized based on the target EEG data to obtain a new deep convolutional neural network model; Performing performance testing on the new deep convolutional neural network model; The channel selection accuracy is determined based on the performance test results.

6. A device for selecting a visual channel for EEG signals for autism, characterized in that: The EEG signal visualization channel selection device for autism comprises: An acquisition module is used to acquire the original EEG data collected by the EEG data acquisition device; Preprocessing module, used to preprocess the raw EEG data; The training module is used to train the deep convolutional neural network model based on the preprocessed EEG data, optimize hyperparameters, and perform performance testing until a deep convolutional neural network model with expected performance is obtained; A generation module is used to obtain, for a certain input sample, a feature map [n × a × b] output by the last convolutional layer of the deep convolutional neural network model that achieves the expected performance, where n is the number of feature maps and a × b is the size of each feature map; calculate Return the gradient information of the feature layer : in, represents the prediction score of the network layer before the softmax layer for category c, is the value of the kth channel of the nth feature map A at position (i, j); Calculate the importance weight of the feature map for category c , where Z represents the scale of the feature layer, which is the product of the number of channels and the bandwidth: Each pair of weights is weighted and linearly combined to form a feature map, and then fed into the ReLU activation function to obtain the class activation map: ； The scoring module is used to process the class activation map of each input sample, accumulate the weights between different frequency bands of each channel, and obtain the total weight of the i-th channel in the n-th image. ; For the i-th channel of the n-th image, if the total weight is higher than the importance threshold T, the importance score of this channel is increased, otherwise no processing is done, and the importance score of each channel is obtained. : ； in, ; The screening module is used to sort the channels in descending order of importance scores and select the channels with the highest preset positions.

7. A device for visualizing EEG signals for autism, characterized in that: The EEG signal visualization channel selection device for autism includes a processor, a memory, and an EEG signal visualization channel selection program for autism stored in the memory and executable by the processor. When the EEG signal visualization channel selection program for autism is executed by the processor, the steps of the EEG signal visualization channel selection method for autism as described in any one of claims 1 to 5 are implemented.

8. A readable storage medium, characterized in that: The readable storage medium stores a program for selecting a channel for visualizing an EEG signal for autism, wherein when the program for visualizing an EEG signal for autism is executed by a processor, the steps of the method for visualizing an EEG signal for autism as described in any one of claims 1 to 5 are implemented.

Images