Lightweight convolutional neural network autism identification method and system based on electroencephalogram data

By combining a lightweight convolutional neural network with the EEGLAB toolkit and multi-scale convolutional layers, the temporal and spatial features of EEG signals are integrated, solving the problem of low efficiency in autism diagnosis in existing models and achieving more efficient autism identification.

CN118986348APending Publication Date: 2024-11-22TIANJIN UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411001548.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing autism diagnostic models based on EEG signals suffer from significant computational waste, resulting in low recognition efficiency and time-consuming training, failing to meet clinical needs.

Method used

Design a lightweight convolutional neural network based on EEG data. Employ multi-scale temporal and spatial convolutional layers, combined with a spatiotemporal attention mechanism, to integrate features from both temporal and spatial dimensions. Preprocess the data using the EEGLAB toolbox, extract features using multi-scale temporal and spatial convolutional kernels, and achieve efficient recognition through separable convolution and classification modules.

Benefits of technology

It achieves more accurate autism classification, reduces the number of parameters and inference time, provides an efficient autism identification scheme, and solves the problem of low efficiency in existing models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118986348A_ABST
    Figure CN118986348A_ABST
Patent Text Reader

Abstract

The invention discloses a lightweight convolutional neural network autism identification method and system based on electroencephalogram data. Comprising the following steps that S1, electroencephalogram data of autistic children and normal healthy children are collected, and an EEGLAB tool box on a Matlab platform is used for conducting preprocessing operation on the electroencephalogram data; s2, adopting a multi-scale time convolution kernel to extract information of different scales of the electroencephalogram signals in the time dimension; s3, integrating the features of the time dimension and the space dimension so as to effectively extract the features of the electroencephalogram signals; s4, extracting information of the electroencephalogram signals in the spatial dimension by adopting a multi-scale spatial convolution kernel; s5, learning and summarizing core features of each feature map, realizing explicit decoupling of relationships in the features and between the features, and combining the learned feature maps together in an optimal mode; and S6, enabling the learned feature map to pass through a full connection layer, and then obtaining a final prediction result through a softmax layer. According to the method, the features of the time dimension and the space dimension are integrated, more accurate autism classification is achieved, a more reliable and efficient solution is provided for application of the autism identification technology, the parameter quantity is smaller, the needed reasoning time is shorter, and the method is suitable for being applied to the autism identification technology. The problems that the autism recognition efficiency of a model is low, and training is time-consuming are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of autism recognition, and in particular to a method and system for recognizing autism using a lightweight convolutional neural network based on electroencephalogram data. Background Art

[0002] Autism is one of the most common neurodevelopmental disorders, which begins in early childhood and lasts a lifetime. Autistic patients are characterized by social interaction disorders, language communication disorders, and repetitive stereotyped behaviors. According to the World Health Organization, the global prevalence of autism is about 1%, and it is increasing worldwide over time. At present, in clinical practice, the diagnosis of autism still relies heavily on the assessment of children's behavior, which can lead to human diagnostic errors. EEG technology records EEG signals generated by the brain through electrodes placed on the scalp, and is widely used due to its excellent temporal resolution, convenience, low setup cost, and non-invasiveness. In order to more accurately identify patients, it is particularly important to combine deep learning with EEG signal data to assist doctors in diagnosis. However, existing models waste a lot of computing power, resulting in inefficient model recognition of autism and time-consuming model training. Therefore, we design a novel spatiotemporal attention mechanism to integrate the features of the temporal and spatial dimensions, and propose a lightweight convolutional neural network method for identifying autism based on EEG data to solve the problem of inefficient model recognition of autism and time-consuming training. Summary of the invention

[0003] The present invention proposes a method for identifying autism based on a lightweight convolutional neural network based on EEG data to solve the problems mentioned in the above background technology. The technical solution steps of the present invention are as follows:

[0004] Step 1: EEG data acquisition and preprocessing module, collect EEG data of autistic children and normal healthy children, and use the EEGLAB toolbox on the Matlab platform to preprocess the EEG data:

[0005] Step 1.1: Use a 128-electrode array to collect EEG data with a sampling rate of 1000 Hz. A total of 73 children's EEG data were obtained, including 41 children with autism aged 4-7 years old and 32 normal healthy children aged 6-8 years old. The EEG data of each child was divided into 10-13 segments, and the length of each segment was 4 seconds;

[0006] Step 1.2: First, bandpass filter the 128-channel EEG data in the range of 1-45Hz; then, use the average reference as a re-reference method; then, detect and process the problematic electrode channels, and repair the bad channels by interpolation; finally, use independent component analysis to remove artifacts. Independent component analysis can decompose multi-channel EEG data into multiple components that are spatially invariant and temporally independent. After generating independent component analysis components, remove the artifact components of each category to obtain a clean EEG; the EEG data of these children can be used as the input of the model: X = [X0, X1, ..., X n ], X n ∈R 128×4000 , where n represents the total number of EEG data segments.

[0007] Step 2: Temporal convolution module, using multi-scale temporal convolution kernels to extract information of EEG signals at different scales in the temporal dimension. The preprocessed EEG signal is sent to the temporal convolution module. First, multi-scale temporal convolution kernels are used to perform one-dimensional convolution operations in parallel. The lengths of these temporal convolution kernels in the temporal dimension are 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the sampling rate, respectively, and their scales are (1, 500), (1, 250), (1, 125), and (1, 62), respectively. Then, the ReLU activation function is used to obtain a richer representation of the features after the convolution operation, and to increase the computing speed and reduce the time cost of model training and inference. Finally, the average pooling method is used to reduce the impact of noise and reduce the feature dimension, thereby obtaining the features of EEG signals at different scales in the temporal dimension:

[0008]

[0009] in, Represents four time convolution kernels of different scales, u = [1, 2, 3, 4], Conv1D represents the time convolution kernel size. The one-dimensional convolution operation is performed, ReLU represents the activation function, and AvgPool represents the average pooling.

[0010] The output of each layer of temporal kernel is concatenated along the feature dimension, and batch normalization is used to stabilize and accelerate the training of the model. T It can be expressed as:

[0011]

[0012] Among them, f BN represents batch normalization, and [·] represents the concatenation of the outputs of the temporal convolution kernels of each layer.

[0013] Step 3: The spatiotemporal attention module integrates the features of the time dimension and the spatial dimension to effectively extract the features of the EEG signal. First, the obtained feature map is subjected to time-domain adaptive average pooling to extract the temporal information; at the same time, all the information in the spatial domain is retained. Then, we use the convolution operation to effectively reduce the number of parameters and further extract the spatial information. Finally, the attention weights are normalized using the softmax function. Since the EEG is very sensitive to amplitude changes, we use scalar multiplication to further expand the output of the softmax function so that it matches it to the same order of magnitude as the features obtained by temporal convolution. Let L be the input feature of the spatiotemporal attention module, then the feature map is generated:

[0014] M(L)=(Softmax(Conv((Pooling(L))) T ))) T *C′ (3)

[0015] Among them, Softmax represents the softmax function, and Pooling represents adaptive average pooling.

[0016] Step 4: Spatial convolution module, using multi-scale spatial convolution kernels to extract information about the EEG signal in the spatial dimension, whose size is related to the position of the EEG channel. First, the multi-scale spatial convolution kernels are subjected to one-dimensional convolution operations in parallel. Here, we use global spatial convolution kernels and hemispherical spatial convolution kernels. On the one hand, the global spatial convolution kernel can capture the global spatial pattern, and its size is the same as the number of channels of the input EEG data, that is, the size of the global spatial convolution kernel is set to (128, 1) and the step size is (1, 1); on the other hand, the hemispherical spatial convolution kernel can effectively capture and model the correlation between the left and right hemispheres, that is, the size and step size of the hemispheric spatial convolution kernel are both set to (64, 1). Then, the ReLU activation function is used to obtain a richer representation of the features after the convolution operation. Finally, the average pooling method is used to reduce the impact of noise and reduce the feature dimension, and the features of the EEG signal at different scales in the spatial dimension are obtained:

[0017]

[0018] in, Represents the global spatial convolution kernel and the hemispherical spatial convolution kernel, v = [0, 1], v = 0 represents the global spatial convolution kernel, v = 1 represents the hemispherical spatial convolution kernel, ConvlD represents the size of the spatial convolution kernel The one-dimensional convolution operation is performed, ReLU represents the activation function, and AvgPool represents the average pooling.

[0019] The output of each layer of spatial convolution kernels is concatenated along the spatial dimension, and batch normalization is used to stabilize and accelerate the training of the model.S It can be expressed as:

[0020]

[0021] Among them, f BN represents batch normalization, and [·] represents the concatenation of the outputs of the spatial convolution kernels of each layer. The output length of the global spatial convolution kernel is 1; the output length of the hemispherical spatial convolution kernel is 2, representing the two hemispheres respectively. After concatenation, the total length is 3.

[0022] Step 5: Separable convolution module, which consists of depthwise convolution and pointwise convolution, where depthwise convolution reduces the number of parameters by sharing a convolution kernel for each input channel, and pointwise convolution reduces the number of parameters by operating each convolution kernel only on the output of one channel. In addition, in depthwise convolution, the core features of each feature map are learned and summarized, achieving explicit decoupling of the relationship between features and features; in pointwise convolution, the learned feature maps are combined in an optimal way.

[0023] Step 6: Classification module is used to classify the feature map learned in step 5. Specifically, the learned feature map is passed through a fully connected layer and then through a softmax layer to obtain the final prediction result.

[0024] A lightweight convolutional neural network system for identifying autism based on EEG data, characterized by comprising an EEG data acquisition and preprocessing module, a temporal convolution module, a spatiotemporal attention module, a spatial convolution module, a separable convolution module and a classification module;

[0025] EEG data acquisition and preprocessing module: collect EEG data of autistic children and normal healthy children, and use the EEGLAB toolbox on the Matlab platform to preprocess the EEG data;

[0026] Temporal convolution module: uses multi-scale temporal convolution kernels to extract information of different scales of EEG signals in the temporal dimension;

[0027] Spatiotemporal attention module: Integrates the features of time and space dimensions to effectively extract EEG signal features;

[0028] Spatial convolution module: uses multi-scale spatial convolution kernels to extract information about EEG signals in the spatial dimension;

[0029] Separable convolution module: The core features of each feature map are learned and summarized, achieving explicit decoupling of the relationship between features and features, and the learned feature maps are combined in an optimal way;

[0030] Classification module: The learned feature map passes through a fully connected layer and then through a softmax layer to obtain the final prediction result.

[0031] The implementation of the method and system for identifying autism based on EEG data using a lightweight convolutional neural network has the following beneficial effects: The present invention proposes a method and system for identifying autism based on EEG data using a lightweight convolutional neural network, which uses multi-scale temporal convolutional layers and spatial convolutional layers to simultaneously learn different feature representations from the temporal dimension and the spatial dimension, introduces a new spatiotemporal attention mechanism, integrates the features of the temporal dimension and the spatial dimension, achieves more accurate autism classification, and provides a more reliable and efficient solution for the application of autism identification technology, and has fewer parameters and requires shorter inference time, solving the problems of low efficiency and time-consuming training in model identification of autism.

[0032] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the specific embodiments. The drawings described below only relate to some embodiments of the present invention, but are not intended to limit the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 A flowchart of a method for identifying autism using a lightweight convolutional neural network based on EEG data according to the present invention;

[0035] Figure 2 A framework diagram of a method for identifying autism using a lightweight convolutional neural network based on EEG data according to the present invention;

[0036] Figure 3 is a spatiotemporal attention framework diagram in an embodiment of the present invention;

[0037] Figure 4 It is the confusion matrix result of the method in the embodiment of the present invention and other control methods.

[0038] Figure 5 A flowchart of a lightweight convolutional neural network system for identifying autism based on EEG data according to the present invention; DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present invention more understandable, the present invention is further described in detail below in combination with specific embodiments and accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] A lightweight convolutional neural network method for identifying autism based on EEG data, such as Figure 1-2 As shown, the following steps are included:

[0041] S1: Collect EEG data of autistic children and normal healthy children, and use the EEGLAB toolbox on the Matlab platform to preprocess the EEG data;

[0042] S1.1: EEG data were collected using a 128-electrode array with a sampling rate of 1000 Hz. EEG data of 73 children were obtained, including 41 children with autism aged 4-7 years old and 32 normal healthy children aged 6-8 years old. The EEG data of each child was divided into 10-13 segments, each of which was 4 seconds long.

[0043] S1.2: First, the 128-channel EEG data in the range of 1-45Hz are bandpass filtered; then, the average reference is used as a re-reference method; then, the problematic electrode channels are detected and processed, and the bad channels are repaired by interpolation; finally, independent component analysis is used to remove artifacts. Independent component analysis can decompose multi-channel EEG data into multiple components that are spatially invariant and temporally independent. After generating independent component analysis components, each type of artifact component is removed to obtain a clean EEG; the EEG data of these children can be used as the input of the model: X = [X0, X1,…, X n ], X n ∈R 128×4000 , where n represents the total number of EEG data segments.

[0044] S2: Temporal convolution module, which uses multi-scale temporal convolution kernels to extract information of EEG signals at different scales in the temporal dimension. The preprocessed EEG signal is sent to the temporal convolution module. First, multi-scale temporal convolution kernels are used to perform one-dimensional convolution operations in parallel. The lengths of these temporal convolution kernels in the temporal dimension are 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the sampling rate, respectively, and their scales are (1, 500), (1, 250), (1, 125), and (1, 62), respectively. Then, the ReLU activation function is used to obtain a richer representation of the features after the convolution operation, and to improve the operation speed and reduce the time cost of model training and inference. Finally, the average pooling method is used to reduce the impact of noise and reduce the feature dimension, thereby obtaining the features of EEG signals at different scales in the temporal dimension:

[0045]

[0046] in, Represents four time convolution kernels of different scales, u = [1, 2, 3, 4], ConvlD represents the time convolution kernel size. The one-dimensional convolution operation is performed, ReLU represents the activation function, and AvgPool represents the average pooling.

[0047] The output of each layer of temporal kernel is concatenated along the feature dimension, and batch normalization is used to stabilize and accelerate the training of the model. T It can be expressed as:

[0048]

[0049] Among them, f BN represents batch normalization, and [·] represents the concatenation of the outputs of the temporal convolution kernels of each layer.

[0050] S3: Spatiotemporal attention module, which integrates the features of time dimension and space dimension to effectively extract EEG signal features. Figure 3 As shown in the figure, first, the obtained feature map is subjected to time-domain adaptive average pooling to extract temporal information; at the same time, all information in the spatial domain is retained. Then, we use convolution operations to effectively reduce the number of parameters and further extract spatial information. Finally, the attention weights are normalized using the softmax function. Since the EEG is very sensitive to amplitude changes, we use scalar multiplication to further expand the output of the softmax function to match it to the same order of magnitude as the features obtained by temporal convolution. Let L be the input feature of the spatiotemporal attention module, then the feature map is generated:

[0051] M(L)=(Softmax(Conv((Poolin(L))) T ))) T*C′ (3)

[0052] Among them, Softmax represents the softmax function, and Pooling represents adaptive average pooling.

[0053] S4: Spatial convolution module, which uses multi-scale spatial convolution kernels to extract information about EEG signals in the spatial dimension, and its size is related to the position of the EEG channel. First, the multi-scale spatial convolution kernels are subjected to one-dimensional convolution operations in parallel. Here, we use global spatial convolution kernels and hemispheric spatial convolution kernels. On the one hand, the global spatial convolution kernel can capture the global spatial pattern, and its size is the same as the number of channels of the input EEG data, that is, the size of the global spatial convolution kernel is set to (128, 1) and the step size is (1, 1); on the other hand, the hemispheric spatial convolution kernel can effectively capture and model the correlation between the left and right hemispheres, that is, the size and step size of the hemispheric spatial convolution kernel are both set to (64, 1). Then, the ReLU activation function is used to obtain a richer representation of the features after the convolution operation. Finally, the average pooling method is used to reduce the impact of noise and reduce the feature dimension, and the features of EEG signals at different scales in the spatial dimension are obtained:

[0054]

[0055] in, Represents the global spatial convolution kernel and the hemispherical spatial convolution kernel, v = [0, 1], v = 0 represents the global spatial convolution kernel, v = 1 represents the hemispherical spatial convolution kernel, ConvlD represents the size of the spatial convolution kernel The one-dimensional convolution operation is performed, ReLU represents the activation function, and AvgPool represents the average pooling.

[0056] The output of each layer of spatial convolution kernels is concatenated along the spatial dimension, and batch normalization is used to stabilize and accelerate the training of the model. S It can be expressed as:

[0057]

[0058] Among them, f BN represents batch normalization, and [·] represents the concatenation of the outputs of the spatial convolution kernels of each layer. The output length of the global spatial convolution kernel is 1; the output length of the hemispherical spatial convolution kernel is 2, representing the two hemispheres respectively. After concatenation, the total length is 3.

[0059] S5: Separable convolution module, which consists of depthwise convolution and pointwise convolution, where depthwise convolution reduces the number of parameters by sharing a convolution kernel for each input channel, and pointwise convolution reduces the number of parameters by operating each convolution kernel only on the output of one channel. In addition, in depthwise convolution, the core features of each feature map are learned and summarized, achieving explicit decoupling of the relationship between features and features; in pointwise convolution, the learned feature maps are combined in an optimal way.

[0060] S6: Classification module, which is used to classify the feature map learned in S5. Specifically, the learned feature map is passed through a fully connected layer and then through a softmax layer to obtain the final prediction result.

[0061] This embodiment sets the following control method:

[0062] Shallow ConvNet: used for EEG signal decoding. It consists of two layers, one for temporal convolution and the other for spatial filtering.

[0063] EEGNet: A compact convolutional neural network architecture designed to decode EEG signals from various brain-computer interface paradigms.

[0064] FBCNet: Aims to effectively extract features from multi-band EEG signals, capture complex EEG signal patterns and perform classification tasks.

[0065] TSception: A multi-scale convolutional neural network architecture designed for emotion recognition based on EEG signals.

[0066] LMDA-Net: is a lightweight neural network model designed to effectively identify EEG signals under different experimental paradigms and different acquisition systems.

[0067] During the experiment, the preprocessed EEG signal data set was evenly divided into five subsets, four of which were used for training and the other for testing. We used five-fold cross validation to evaluate the accuracy of the method of the present invention and the control method, and we compared the parameters and inference time of the method of the present invention and the control method, as shown in Table 1. The confusion matrix results are shown in Figure 4 shown.

[0068] Table 1 Comparison of test results between the method of the present invention and the control method

[0069]

[0070] It can be seen from Table 1 above that the performance of the method of the present invention is obviously better than that of the control method, and the number of parameters and the inference time are also significantly less than those of the control method, which shows that the present invention can effectively use a lightweight convolutional neural network based on EEG data to predict whether the subject child suffers from autism.

[0071] A lightweight convolutional neural network system for identifying autism based on EEG data, Figure 5 As shown, the following steps are included:

[0072] E1, EEG data acquisition and preprocessing module, collects EEG data of autistic children and normal healthy children, and uses the EEGLAB toolbox on the Matlab platform to preprocess the EEG data;

[0073] E2, temporal convolution module, uses multi-scale temporal convolution kernels to extract information of different scales of EEG signals in the temporal dimension;

[0074] E3, spatiotemporal attention module, integrates the features of time and space dimensions to effectively extract EEG signal features;

[0075] E4, spatial convolution module, uses multi-scale spatial convolution kernels to extract information of EEG signals in the spatial dimension;

[0076] E5, separable convolution module, the core features of each feature map are learned and summarized, achieving explicit decoupling of the relationship between features and features, and the learned feature maps are combined in an optimal way;

[0077] E6, classification module, passes the learned feature map through a fully connected layer and then through the softmax layer to obtain the final prediction result.

[0078] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for identifying autism based on a lightweight convolutional neural network based on EEG data, characterized in that: The specific steps are as follows: Step 1: EEG data acquisition and preprocessing module, collect EEG data of autistic children and normal healthy children, and use the EEGLAB toolbox on the Matlab platform to preprocess the EEG data: Step 2: Temporal convolution module, using multi-scale temporal convolution kernels to extract information of EEG signals at different scales in the temporal dimension. The preprocessed EEG signal is sent to the temporal convolution module. First, multi-scale temporal convolution kernels are used to perform one-dimensional convolution operations in parallel. The lengths of these temporal convolution kernels in the temporal dimension are 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the sampling rate, respectively, and their scales are (1, 500), (1, 250), (1, 125), and (1, 62), respectively. Then, the ReLU activation function is used to obtain a richer representation of the features after the convolution operation, and to increase the computing speed and reduce the time cost of model training and inference. Finally, the average pooling method is used to reduce the impact of noise and reduce the feature dimension, thereby obtaining the features of EEG signals at different scales in the temporal dimension: in, Represents four time convolution kernels of different scales, u = [1, 2, 3, 4], Conv1D represents the time convolution kernel size. The one-dimensional convolution operation is performed, ReLU represents the activation function, and AvgPool represents the average pooling. The output of each layer of temporal kernel is concatenated along the feature dimension, and batch normalization is used to stabilize and accelerate the training of the model. T It can be expressed as: Among them, f BN represents batch normalization, and [·] represents the concatenation of the outputs of the temporal convolution kernels of each layer. Step 3: The spatiotemporal attention module integrates the features of the time dimension and the spatial dimension to effectively extract the features of the EEG signal. First, the obtained feature map is subjected to time-domain adaptive average pooling to extract the temporal information; at the same time, all the information in the spatial domain is retained. Then, we use the convolution operation to effectively reduce the number of parameters and further extract the spatial information. Finally, the attention weights are normalized using the softmax function. Since the EEG is very sensitive to amplitude changes, we use scalar multiplication to further expand the output of the softmax function so that it matches it to the same order of magnitude as the features obtained by temporal convolution. Let L be the input feature of the spatiotemporal attention module, then the feature map is generated: M(L)=(Softmax(Conv((Pooling(L)) T ))) T *C′ (3) Among them, Softmax represents the softmax function, and Pooling represents adaptive average pooling. Step 4: Spatial convolution module, using multi-scale spatial convolution kernels to extract information about the EEG signal in the spatial dimension, whose size is related to the position of the EEG channel. First, the multi-scale spatial convolution kernels are subjected to one-dimensional convolution operations in parallel. Here, we use global spatial convolution kernels and hemispherical spatial convolution kernels. On the one hand, the global spatial convolution kernel can capture the global spatial pattern, and its size is the same as the number of channels of the input EEG data, that is, the size of the global spatial convolution kernel is set to (128, 1) and the step size is (1, 1); on the other hand, the hemispherical spatial convolution kernel can effectively capture and model the correlation between the left and right hemispheres, that is, the size and step size of the hemispheric spatial convolution kernel are both set to (64, 1). Then, the ReLU activation function is used to obtain a richer representation of the features after the convolution operation. Finally, the average pooling method is used to reduce the impact of noise and reduce the feature dimension, and the features of the EEG signal at different scales in the spatial dimension are obtained: in, Represents the global spatial convolution kernel and the hemispherical spatial convolution kernel, v = [0, 1], v = 0 represents the global spatial convolution kernel, v = 1 represents the hemispherical spatial convolution kernel, Conv1D represents the size of the spatial convolution kernel The one-dimensional convolution operation is performed, ReLU represents the activation function, and AvgPool represents the average pooling. The output of each layer of spatial convolution kernels is concatenated along the spatial dimension, and batch normalization is used to stabilize and accelerate the training of the model. S It can be expressed as: Among them, f BN represents batch normalization, and [·] represents the concatenation of the outputs of the spatial convolution kernels of each layer. The output length of the global spatial convolution kernel is 1; the output length of the hemispherical spatial convolution kernel is 2, representing the two hemispheres respectively. After concatenation, the total length is 3. Step 5: Separable convolution module, which consists of depthwise convolution and pointwise convolution, where depthwise convolution reduces the number of parameters by sharing a convolution kernel for each input channel, and pointwise convolution reduces the number of parameters by operating each convolution kernel only on the output of one channel. In addition, in depthwise convolution, the core features of each feature map are learned and summarized, achieving explicit decoupling of the relationship between features and features; in pointwise convolution, the learned feature maps are combined in an optimal way. Step 6: Classification module is used to classify the feature map learned in step 5. Specifically, the learned feature map is passed through a fully connected layer and then through a softmax layer to obtain the final prediction result.

2. According to claim 1, a method for identifying autism using a lightweight convolutional neural network based on EEG data, characterized in that: The sub-steps of the EEG data acquisition and preprocessing module in step 1 are: Step 1.1: Use a 128-electrode array to collect EEG data with a sampling rate of 1000 Hz. A total of 73 children's EEG data were obtained, including 41 children with autism aged 4-7 years old and 32 normal healthy children aged 6-8 years old. The EEG data of each child was divided into 10-13 segments, and the length of each segment was 4 seconds; Step 1.2: First, bandpass filter the 128-channel EEG data in the range of 1-45Hz; then, use the average reference as a re-reference method; then, detect and process the problematic electrode channels, and repair the bad channels by interpolation; finally, use independent component analysis to remove artifacts. Independent component analysis can decompose multi-channel EEG data into multiple components that are spatially invariant and temporally independent. After generating independent component analysis components, remove the artifact components of each category to obtain a clean EEG; the EEG data of these children can be used as the input of the model: X = [X0, X1, ..., X n ], X n ∈R 128x4000 , where n represents the total number of EEG data segments.

3. A lightweight convolutional neural network system for identifying autism based on EEG data, characterized in that: It includes EEG data acquisition and preprocessing module, temporal convolution module, spatiotemporal attention module, spatial convolution module, separable convolution module, and classification module; EEG data acquisition and preprocessing module: collect EEG data of autistic children and normal healthy children, and use the EEGLAB toolbox on the Matlab platform to preprocess the EEG data; Temporal convolution module: uses multi-scale temporal convolution kernels to extract information of different scales of EEG signals in the temporal dimension; Spatiotemporal attention module: Integrates the features of time and space dimensions to effectively extract EEG signal features; Spatial convolution module: uses multi-scale spatial convolution kernels to extract information about EEG signals in the spatial dimension; Separable convolution module: The core features of each feature map are learned and summarized, achieving explicit decoupling of the relationship between features and features, and the learned feature maps are combined in an optimal way; Classification module: The learned feature map passes through a fully connected layer and then through a softmax layer to obtain the final prediction result.

Citation Information

Cited By

  • Frequency prior feature guided lightweight motor imagery classification method and device

    CN119884889A

  • Training method and training device for brain cognitive state classification model

    CN120724268A

  • Autistic child evaluation system and device based on electroencephalogram large model

    CN121313116A

  • Autism Assessment System and Device Based on Electroencephalogram (EEG) Model

    CN121313116B