ADHD classification method based on electroencephalogram signals and channel selection attention mechanism

Abstract

The invention discloses an ADHD classification method based on electroencephalogram signals and a channel selection attention mechanism, and particularly relates to the technical field of electroencephalogram signal classification, and the method comprises the steps: obtaining the public electroencephalogram data of ADHD children, ADD children and HC children in task states, carrying out the preprocessing of the data, carrying out the analysis of independent components, separating data components, removing artifact components, and carrying out the classification of the ADHD children, the ADD children and the HC children; the method comprises the steps of obtaining pure electroencephalogram signals, segmenting data and dividing a data set, designing an ADHD electroencephalogram signal classification algorithm based on a multi-scale channel selective attention mechanism and a neural network, capturing features of different scales through a multi-scale CNN, focusing on correlation enhancement features among channels by using the channel selective attention mechanism, and classifying the electroencephalogram signals to obtain an ADHD electroencephalogram signal classification algorithm based on the multi-scale channel selective attention mechanism. And finally, capturing a long-distance dependency relationship through a self-attention mechanism and fusing global features.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal classification, and in particular to an ADHD classification method based on EEG signals and a channel selection attention mechanism. Background Art

[0002] At present, a variety of ADHD diagnostic assistance methods based on deep learning have been studied. The invention patent with publication number CN112155549A discloses an ADHD disease diagnosis auxiliary decision system based on a deep convolutional pulse neural network. The signal-to-noise ratio feature selection method is used to preprocess the functional magnetic resonance imaging data, and the forward algorithm is used to encode it into a pulse sequence, and finally the deep convolutional pulse neural network is used for diagnosis prediction. The system does not require manual feature annotation, and optimizes the prediction model by assigning different adjustment weights to medical records, thereby improving the reliability of diagnostic prediction results. However, the system still has some potential shortcomings: it relies on high-quality functional magnetic resonance imaging data, and the technology has problems such as high acquisition cost and long time consumption, which may limit the promotion and application of the system. In addition, the model is complex and has limited interpretability, and the generalization ability and potential risks also need further research and evaluation; The invention patent with the publication number CN117496396A discloses a method and system for ADHD multi-feature extraction and fusion classification based on original videos. By analyzing the facial movements, eye movements and head movements of the subjects when watching the video, multimodal features are extracted and fused, and finally classified using a deep learning network. This method has the advantages of being non-invasive, low-cost, easy to promote, high accuracy, and a wide range of applications. However, this method also has some potential shortcomings: the evaluation results are easily affected by environmental factors, the multimodal feature extraction process may be more complicated, the training of the deep learning model requires a large amount of video data, and the generalization ability of the model in different populations and different scenarios needs to be further verified; The invention patent with publication number CN111528859A discloses a children's ADHD screening and evaluation system based on multimodal deep learning technology, which combines multimodal information such as ADHD diagnostic psychology scale, eye movement attention, expression, posture, and mouse and keyboard operation, and uses the BERT model for time series information fusion, and finally realizes the screening and evaluation of children's ADHD. The system has the advantages of high efficiency and high accuracy. However, the system also has some potential shortcomings: it relies on a large amount of multimodal data for training, and the cost of data collection and annotation is high. In addition, the model has poor interpretability, and it is difficult to clearly point out which specific factors lead to the diagnosis of ADHD, which may limit its clinical value in practical applications.

[0003] The above-mentioned prior art shows that deep learning has great potential in assisting the diagnosis of ADHD, but it also faces challenges such as data dependence and poor interpretability. In the future, more research is needed to explore more reliable, easy-to-use and interpretable deep learning diagnostic methods for ADHD. As a non-invasive neurophysiological technology with high temporal resolution, real-time monitoring and low price, EEG signals show unique advantages in ADHD diagnostic research. Studies have shown that there are significant differences in EEG between ADHD patients and healthy people. However, most of the existing EEG-based ADHD diagnostic models ignore the relevant information between channels, and using this information can improve the feature expression ability, generalization ability and robustness of the model, help reduce the risk of overfitting, and improve model efficiency and performance. Therefore, the present invention proposes an ADHD classification method based on EEG signals and channel selection attention mechanism, which further improves the accuracy of ADHD diagnosis by capturing information between different channels. Summary of the invention

[0004] To this end, the present invention provides an ADHD classification method based on EEG signals and channel selection attention mechanism to solve the problems raised in the background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solutions: The present invention discloses an ADHD classification method based on EEG signals and a channel selective attention mechanism. Public EEG data of ADHD children, ADD (a subtype of ADHD) children and healthy (HC) children in a task state are obtained and preprocessed. Independent component analysis (ICA) is used to separate data components, and artifact components are removed by expert visual inspection to obtain pure EEG signals and segment the data and divide the data sets. The channel selective attention mechanism is used to focus on the correlation enhancement features between channels. An ADHD EEG signal classification algorithm based on a multi-scale channel selective attention mechanism and a neural network is designed. Features of different scales are captured through a multi-scale CNN. Finally, a self-attention mechanism is used to capture long-distance dependencies and fuse global features, thereby generating richer and more accurate feature representations for improving the performance of the model in classification tasks.

[0006] Preferably, data preprocessing includes removing useless electrodes, re-referencing, filtering, segmenting, and eliminating useless data segments.

[0007] The channel selective attention mechanism is designed as follows: The input of the attention mechanism is output by two convolutional kernel branches of different scales; assuming that the feature map and , where 1, W, and C represent the number of EEG leads, EEG time points, and feature channels after the pre-neural network transformation, respectively. They come from two branches of different scales, where W1 and C1 both represent the feature map dimensions obtained from the output of the pre-neural network branch 1, and W2 and C2 both represent the feature map dimensions obtained from the output of the pre-neural network branch 2; First, and Add point by point to obtain a multi-scale fusion feature map ; where W and C superscripts refer to the dimensions of the new feature map obtained by adding the pre-neural network branch 1 and the pre-neural network branch 2, respectively. Then, Use global average pooling GAP and global maximum pooling GMP respectively to obtain statistics on the channel dimension and ;then, The interaction between channels is learned through a bottleneck block consisting of two fully connected layers FC; The first FC is followed by a batch normalization (BN) and an ELU activation function to compress the number of channels to 1 / 4 of the input and update the value. The second FC is used to expand the number of channels to twice the input. , for subsequent allocation to and ; The interaction between channels is learned through a bottleneck block consisting of two fully connected layers FC; and Obtained by point-by-point addition ;Will Before The characteristic value of each channel After being activated by the softmax function, they are distributed to ,Will After The characteristic value of each channel After being activated by the softmax function, they are distributed to , and get the final output of SCAM and ,accomplish Input feature maps assigned to different scales and The original input and final output data have the same shape, making the SCAM module versatile without losing its effectiveness.

[0008] Preferably, the steps for implementing the ADHD EEG signal classification algorithm based on the multi-scale channel selective attention mechanism and neural network include: S4-1, randomly initialize the weight parameter W and bias parameter b of the neural network, and set hyperparameters such as learning rate lr, batch_size of data loaded per batch, and epoch of training. The loss function uses cross entropy, and the Adam optimizer is used to update the model parameters; S4-2. Input the divided training set into the initialized neural network model, and use (x, y) to represent the EEG signals in the training set and their corresponding ADHD, ADD and HC category labels; where vector x represents the EEG signal data; vector y represents the label; the dimension of vector x is , where H and W represent the number of channels and time points of the EEG signal, respectively; S4-3. Design model ,Model It can classify the input x into three categories: ADHD, ADD and HC; Model The following architectures are included: (1) Input layer: First, the preprocessed data is standardized as follows:

[0009] Among them, X represents the original data, μ represents the mean, and σ represents the standard deviation; After normalization, the dimension of the given input data is , where H and W represent the number of EEG channels and time points respectively; a feature channel dimension is added to the last dimension, so that the input data is ; (2) Local feature extraction layer: Given the dimension of input data is , first use a convolution operation with a filter of size (1, 32) to keep the dimension unchanged; and then use the same padding to get ; Then batch normalization is performed and the ELU activation function is applied to introduce nonlinearity into the network; (3) Channel feature fusion layer: Given the dimension of input data is , where a depthwise convolution is applied with a kernel size of (H, 1) and a depth multiplier parameter of 2, resulting in a data dimension of ; Subsequently, batch normalization is performed along the last axis of the tensor and the ELU activation function is applied to introduce nonlinearity into the network; (4) SCA multi-scale temporal feature extraction layer: Given the dimension of input data is , average pooling is applied along the time dimension and the input tensor is downsampled; then dropout regularization with a dropout rate of 0.5 is applied to transform the data dimension to ; Next, a dual-branch convolution block is used to extract high-level temporal features, using convolution kernels of (1*3) and (1*5) sizes; the convolution block function is repeatedly applied, including convolution layers, batch normalization, and activation functions, and the dimension of the data does not change in this operation; The SCA mechanism is introduced and applied; after that, the extracted features are subjected to another round of average pooling and discard regularization. At this time, the output data is the two branches ; Then, continue to use convolution blocks to capture features; Finally, the output features of the two branches are fused together again using the SCA mechanism to promote the integration of multi-scale temporal information; (5) Self-attention global feature fusion layer: Given two input data with dimensions of , , concatenate the features of the two branches along the channel axis to create a fused feature tensor ; Next, a convolution block with a kernel size of (1*1) and a number of 2C is applied to the fused features to further refine and extract discriminative information, followed by batch normalization and activation function; Then, delete the first dimension in the data, and the data dimension is ; Then the Self-attention mechanism is applied to capture the dependencies between global features; After the Self-attention operation, the feature tensor is flattened to ensure that the feature dimension is suitable for input into the fully connected layer; Finally, a fully connected layer with softmax activation function is applied to perform classification based on the learned features; S4-4. Use the Adam optimizer to update the weight parameter W and bias parameter b of the neural network model, and continuously update the training until the epoch training is completed.

[0010] The present invention verifies the effectiveness of the algorithm according to the evaluation index, and adopts three objective evaluation indexes to verify the performance of the present invention in the classification task, and obtains the accuracy, sensitivity and specificity.

[0011]

[0012]

[0013]

[0014] Among them, TP: true positive, FP: false positive, FN: false negative, TN: true negative.

[0015] The present invention has the following advantages: 1. By introducing deep convolution and point convolution to extract features from EEG signals, and adopting a multi-scale, dual-branch network structure, the morphological features of EEG signals can be effectively extracted.

[0016] 2. A classification method for ADHD based on EEG signals and channel selection attention mechanism is proposed. By adopting a multi-scale, dual-branch network structure and combining the self-attention mechanism to build a model, the global correlation is effectively captured, thereby generating richer and more accurate feature representations to improve the performance of the model in classification tasks. The accuracy of ADHD diagnosis is further improved by capturing information between different channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall network structure diagram adopted by the present invention; Figure 2 It is a structural diagram of the SCAM module of the present invention. DETAILED DESCRIPTION

[0018] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. 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.

[0019] This embodiment refers to Figure 1 and Figure 2 .

[0020] Step 1: Data Collection This example uses an open source dataset from the Technical University of Dresden, Germany (which can be viewed at https: / / osf.io / 6594x / ) as the research subject. In this study, a total of 144 children were included, including 44 healthy controls (HC groups), 52 ADHD patients, and 48 ADD patients. Informed consent was obtained from all participants and their guardians in accordance with the Declaration of Helsinki. The diagnosis and classification of these children were completed by psychiatrists and psychologists based on standard clinical indicators, family and school interviews, questionnaires, and IQ and attention tests. There were no significant differences in age, IQ, and gender between the groups.

[0021] In the task of this dataset, participants were asked to press a button at an estimated 1200 ms after seeing a white square on a black background on the display. If the reaction time was between 1000 and 1400 ms, it was considered an accurate response. Participants received visual feedback to let them know whether their response was correct. For correct responses, a green "smiley face" appeared on the screen; if the reaction time exceeded the required time window, a red "frown" and a "too early" or "too late" prompt appeared on the screen.

[0022] Step 2: Data Processing Step 2-1: The raw data was subjected to independent component analysis and a series of EEG preprocessing and denoising to obtain clean data, which was divided into 33,902 segments, including 10,742 items for the ADHD group, 13,031 items for the ADD group, and 10,129 items for the control group (HC group). Each segment was 1.5 seconds long, with a sampling rate of 256 Hz and 56 channels.

[0023] Step 2-2: This study adopted a five-fold cross-validation method. The laboratory used all 33,902 data fragments, of which 6,000 fragments were divided into test sets, and the remaining 27,902 were evenly divided into 5 parts. Each time, 4 parts were selected as training sets and 1 part was selected as a validation set.

[0024] Step 3: ADHD EEG signal classification algorithm based on multi-scale channel selective attention mechanism and neural network: Step 3-1: Randomly initialize the weight parameter W and bias parameter b of the neural network, and set hyperparameters such as learning rate lr, batch_size of data loaded per batch, and epoch of training. The loss function uses cross entropy and the Adam optimizer is used to update the model parameters.

[0025] Step 3-2: Input the training set divided in step 2-2 into the neural network model, and use (x, y) to represent the EEG signals in the training set and their corresponding ADHD, ADD and HC category labels.

[0026] Step 3-3: Vector x, which represents the EEG signal data, the dimension of vector x is , where H and W represent the number of channels and time points of the EEG signal respectively; the design function ,function It can classify the input x into three categories: ADHD, ADD and HC.

[0027] Model The following architectures are included: (1) Input layer: First, the preprocessed data is standardized as follows: ; Among them, X represents the original data, μ represents the mean, and σ represents the standard deviation; After normalization, the dimension of the given input data is , where H and W represent the number of EEG channels and time points respectively; a feature channel dimension is added to the last dimension, so that the input data is ; (2) Local feature extraction layer: Given the dimension of input data is , first use a convolution operation with a filter of size (1, 32) to keep the dimension unchanged; and then use the same padding to get ; Then batch normalization is performed and the ELU activation function is applied to introduce nonlinearity into the network; (3) Channel feature fusion layer: Given the dimension of input data is , where a depthwise convolution is applied with a kernel size of (H, 1) and a depth multiplier parameter of 2, resulting in a data dimension of ; Subsequently, batch normalization is performed along the last axis of the tensor and the ELU activation function is applied to introduce nonlinearity into the network; (4) SCA multi-scale temporal feature extraction layer: Given the dimension of input data is , average pooling is applied along the time dimension and the input tensor is downsampled; then dropout regularization with a dropout rate of 0.5 is applied to transform the data dimension to ; Next, a dual-branch convolution block is used to extract high-level temporal features, using convolution kernels of (1*3) and (1*5) sizes; the convolution block function is repeatedly applied, including convolution layers, batch normalization, and activation functions, and the dimension of the data does not change in this operation; The SCA mechanism is introduced and applied; after that, the extracted features are subjected to another round of average pooling and discard regularization. At this time, the output data is the two branches ; Then, continue to use convolution blocks to capture features; Finally, the output features of the two branches are fused together again using the SCA mechanism to promote the integration of multi-scale temporal information; (5) Self-attention global feature fusion layer: Given two input data with dimensions of , , concatenate the features of the two branches along the channel axis to create a fused feature tensor ; Next, a convolution block with a kernel size of (1*1) and a number of 2C is applied to the fused features to further refine and extract discriminative information, followed by batch normalization and activation function; Then, delete the first dimension in the data, and the data dimension is ; Then the Self-attention mechanism is applied to capture the dependencies between global features; After the Self-attention operation, the feature tensor is flattened to ensure that the feature dimension is suitable for input into the fully connected layer; Finally, a fully connected layer with softmax activation function is applied to perform classification based on the learned features; In the neural network proposed above, CNN, the independently proposed channel selection attention mechanism (Selective Channels Attention Mechanism, SCAM) and self-attention mechanism (Self-attention) are combined. The specific scheme uses multi-scale CNN to capture features of different scales, uses SCAM to focus on the correlation between channels to enhance features, and finally uses Self-attention to capture long-distance dependencies and fuse global features. It is mainly divided into five modules: input layer, local feature extraction layer, channel feature fusion layer, SCA multi-scale temporal feature extraction layer, and global feature fusion layer. Figure 1 shown.

[0028] The above-mentioned SCAM is designed by the present invention, and its input is output by two convolution kernel branches of different scales.

[0029] Assuming that the feature map is given and , which come from two branches of different scales. First, we and Add point by point to obtain a multi-scale fusion feature map Then, Use global average pooling (GAP) and global maximum pooling (GMP) to obtain statistics on the channel dimension and .

[0030] then, The interaction between channels is learned through a bottleneck block consisting of two fully connected layers (FC). The first FC is followed by batch normalization (BN) and ELU activation function to compress the number of channels to 1 / 4 of the input and update the value, and the second FC is used to expand the number of channels to twice the input to obtain , for subsequent allocation to and .right Perform the same operation as above to obtain . Further, and Obtained by point-by-point addition In order to Input feature maps assigned to different scales and ,we will Before The characteristic value of each channel After being activated by the softmax function, they are distributed to ,Will After The characteristic value of each channel After being activated by the softmax function, they are distributed to , and get the final output of SCAM and The original input and final output data have the same shape, making the SCAM module versatile without losing its effectiveness.

[0031] This module is used to pay attention to the correlation between channel features, and the self-attention mechanism is used to connect the context information to generate richer and more accurate feature representations, which is used to improve the performance of the model in classification tasks.

[0032] Step 3-4: Use the Adam optimizer to update the weight parameter W and bias parameter b of the neural network. Continue to update the training until the epoch training is completed.

[0033] Step 4: Verify the effectiveness of the algorithm based on the evaluation indicators.

[0034] This embodiment uses three objective evaluation indicators to verify the performance of the present invention in classification tasks, namely, accuracy, sensitivity, and specificity.

[0035]

[0036]

[0037]

[0038] Among them, TP: true positive, FP: false positive, FN: false negative, TN: true negative.

[0039] In this embodiment, the designed ADHD EEG signal classification algorithm based on multi-scale channel selective attention mechanism and neural network is trained and verified on data with the existing EEG classification models ShallowConvNet, DeepConvNet, EEGNet, and CNN-LSTM to compare performance, as shown in Table 1.

[0040] Table 1. Performance comparison of deep learning models

[0041] SCANet achieved the best performance in most evaluation indicators, with accuracy reaching 99.22%, sensitivity reaching 99.23%, and specificity reaching 99.77%. Compared with CNN-LSTM, SCANet improved accuracy, specificity, and sensitivity by 0.99%, 0.99%, and 0.66%, respectively.

[0042] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. ADHD classification method based on EEG signals and channel selection attention mechanism, obtains and preprocesses the public EEG data of ADHD children, ADD children and HC children in task state, uses independent component analysis to separate data components, removes artifact components, obtains pure EEG signals, segments the data and divides the data set, which is characterized by: The channel selective attention mechanism is used to focus on the correlation enhancement features between channels. An ADHD EEG signal classification algorithm based on multi-scale channel selective attention mechanism and neural network is designed. Features of different scales are captured through multi-scale CNN. Finally, the long-distance dependencies are captured and the global features are integrated through the self-attention mechanism.

2. The ADHD classification method based on EEG signals and channel selection attention mechanism according to claim 1, characterized in that: Data preprocessing includes removing useless electrodes, re-referencing, filtering, segmentation, and eliminating useless data segments.

3. The ADHD classification method based on EEG signals and channel selection attention mechanism according to claim 1, characterized in that: The channel selective attention mechanism is designed as follows: The input of the attention mechanism is output by two convolutional kernel branches of different scales; Assuming that the feature map is given and , where 1, W, and C represent the number of EEG leads, EEG time points, and feature channels after the pre-neural network transformation, respectively. They come from two branches of different scales, where W1 and C1 both represent the feature map dimensions obtained from the output of the pre-neural network branch 1, and W2 and C2 both represent the feature map dimensions obtained from the output of the pre-neural network branch 2; First, and Add point by point to obtain a multi-scale fusion feature map ; where W and C superscripts refer to the dimensions of the new feature map obtained by adding the pre-neural network branch 1 and the pre-neural network branch 2, respectively. Then, Use global average pooling GAP and global maximum pooling GMP respectively to obtain statistics on the channel dimension and ;then, The interaction between channels is learned through a bottleneck block consisting of two fully connected layers FC; The first FC is followed by a batch normalization (BN) and an ELU activation function to compress the number of channels to 1 / 4 of the input and update the value. The second FC is used to expand the number of channels to twice the input. , for subsequent allocation to and ; The interaction between channels is learned through a bottleneck block consisting of two fully connected layers FC; and Obtained by point-by-point addition ;Will Before The characteristic value of the channel After being activated by the softmax function, they are distributed to ,Will After The characteristic value of the channel After being activated by the softmax function, they are distributed to , and get the final output of SCAM and ,accomplish Input feature maps assigned to different scales and .

4. The ADHD classification method based on EEG signals and channel selection attention mechanism according to claim 1, characterized in that: The implementation steps of the ADHD EEG signal classification algorithm based on multi-scale channel selective attention mechanism and neural network include: S4-1, randomly initialize the weight parameter W and bias parameter b of the neural network, set the learning rate lr, the amount of data loaded in each batch batch_size and the training round epoch, use the cross entropy loss function, and use the Adam optimizer to update the model parameters; S4-2. Input the divided training set into the initialized neural network model, and use (x, y) to represent the EEG signals in the training set and their corresponding ADHD, ADD and HC category labels; where vector x represents the EEG signal data; vector y represents the label; the dimension of vector x is , where H and W represent the number of channels and time points of the EEG signal, respectively; S4-3. Design model ,Model It can classify the input x into three categories: ADHD, ADD and HC; Model The following architectures are included: (1) Input layer: First, the preprocessed data is standardized as follows: ； Among them, X represents the original data, μ represents the mean, and σ represents the standard deviation; After normalization, the dimension of the given input data is , where H and W represent the number of EEG channels and time points respectively; a feature channel dimension is added to the last dimension, so that the input data is ; (2) Local feature extraction layer: Given the dimension of input data is , first use a convolution operation with a filter of size (1, 32) to keep the dimension unchanged; and then use the same padding to get ; Then batch normalization is performed and the ELU activation function is applied to introduce nonlinearity into the network; (3) Channel feature fusion layer: Given the dimension of input data is , where a depthwise convolution is applied with a kernel size of (H, 1) and a depth multiplier parameter of 2, resulting in a data dimension of ; Subsequently, batch normalization is performed along the last axis of the tensor and the ELU activation function is applied to introduce nonlinearity into the network; (4) SCA multi-scale temporal feature extraction layer: Given the dimension of input data is , average pooling is applied along the time dimension and the input tensor is downsampled; then dropout regularization with a dropout rate of 0.5 is applied to transform the data dimension to ; Next, a dual-branch convolution block is used to extract high-level temporal features, using convolution kernels of (1*3) and (1*5) sizes; the convolution block function is repeatedly applied, including convolution layers, batch normalization, and activation functions, and the dimension of the data does not change in this operation; The SCA mechanism is introduced and applied; after that, the extracted features are subjected to another round of average pooling and discard regularization. At this time, the output data is the two branches ; Then, continue to use convolution blocks to capture features; Finally, the output features of the two branches are fused together again using the SCA mechanism to promote the integration of multi-scale temporal information; (5) Self-attention global feature fusion layer: Given two input data with dimensions of , , concatenate the features of the two branches along the channel axis to create a fused feature tensor ; Next, a convolution block with a kernel size of (1*1) and a number of 2C is applied to the fused features to further refine and extract discriminative information, followed by batch normalization and activation function; Then, delete the first dimension in the data, and the data dimension is ; Then the Self-attention mechanism is applied to capture the dependencies between global features; After the Self-attention operation, the feature tensor is flattened to ensure that the feature dimension is suitable for input into the fully connected layer; Finally, a fully connected layer with softmax activation function is applied to perform classification based on the learned features; S4-4. Use the Adam optimizer to update the weight parameter W and bias parameter b of the neural network model, and continuously update the training until the epoch training is completed.

Images