Children's sleep staging method based on collecting EEG signals with headband electrodes

Abstract

The present invention relates to a child sleep staging method based on collecting electroencephalogram signals by headband electrodes, belongs to the technical field of sleep staging, solves the problem of poor comfort and coordination when traditional polysomnography monitoring equipment is used on children, and includes the steps of: constructing a Web client and a server; the headband electrode equipment collects electroencephalogram signals of multiple child subjects; the server adopts a new wavelet threshold denoising strategy based on EMD to denoise the electroencephalogram signals to obtain sleep electroencephalogram data, performs data segmentation and elimination on the data, constructs a time-frequency feature map, trains and tests a multi-view hybrid neural network model, obtains a child sleep staging model, uses the model to perform child sleep staging, and displays the child sleep staging results by the Web client. The present invention can improve the comfort and coordination of children in the data collection process, improve signal quality, and effectively improve the classification accuracy and model performance of the child sleep staging model.

Description

Technical Field

[0001] The invention belongs to the technical field of sleep staging, and in particular relates to a method for children's sleep staging based on collecting electroencephalogram signals using headband electrodes. Background Art

[0002] Sleep has an important impact on the human immune system, metabolism, and brain memory. Good sleep is essential for the physical and mental health and intellectual development of children during their growth and development period. However, sleep disorders not only damage children's cognitive abilities such as learning and memory, but may also cause a series of long-term complications. Sleep staging helps medical staff monitor the sleep quality of children and helps child patients to detect and diagnose early. However, sleep staging clinically requires manual judgment by experts combined with polysomnography data, which is extremely time-consuming and highly dependent on the expert's experience. Therefore, it is of great significance to study methods that can efficiently and accurately achieve automatic sleep staging.

[0003] Deep learning methods can automatically learn the feature representation of data, reduce the burden of feature engineering, and have better feature learning capabilities than traditional machine learning methods. At present, deep learning technology is widely used in automatic sleep staging tasks, but most of them are based on adults, and there are few studies on sleep staging for children. Compared with adults, children's sleep patterns are affected by age factors and differ from adults in some stages. At the same time, the current children's sleep staging model usually only uses single-dimensional information for feature extraction, which may lead to insufficient feature extraction, thereby limiting the performance of the sleep staging model. Therefore, the adult sleep staging method is not completely applicable to children, and there is an urgent need to conduct more in-depth research on the children's sleep staging method based on the children's sleep structure. In addition, the existing sleep data collection mainly uses polysomnography (PSG) equipment, but this equipment may have some problems. For example, the use of it on children may lead to poor comfort and cooperation, which greatly affects the children's sleep quality, and then affects the quality of sleep EEG signals. Considering the comfort and cooperation of children during the collection or monitoring process, there is an urgent need for a convenient and comfortable sleep collection device. Summary of the invention

[0004] The purpose of the present invention is to provide a children's sleep staging method based on headband electrodes to collect EEG signals, which mainly solves the problems of poor comfort and coordination when traditional polysomnography (PSG) equipment is used on children, and the problem that traditional adult sleep staging methods are not suitable for children due to the differences in children's and adults' sleep patterns.

[0005] To solve the above-mentioned deficiencies in the prior art, the present invention adopts the following technical solutions:

[0006] A method for children's sleep staging based on collecting EEG signals with headband electrodes, the method comprising the following steps:

[0007] Step 1: Build the web client and server;

[0008] Step 2: The headband electrode device collects EEG signals of multiple child subjects during sleep, and sends the EEG signals to the server through the Web client;

[0009] Step 3: The server performs the following steps:

[0010] Step 3.1: Using a new wavelet threshold denoising strategy based on EMD to pre-process the EEG signal to obtain sleep EEG data;

[0011] Step 3.2: segmenting and eliminating the sleep EEG data to construct a time-frequency feature map;

[0012] Step 3.3: Construct a multi-view hybrid neural network model, which includes a time-domain dual-branch convolution module, a time-frequency feature map extraction module, a fusion module and a classification module. The time-domain dual-branch convolution module takes the time domain signal as input, and includes two parallel convolution network branches with different kernel sizes. The first convolution network branch includes sequentially connected convolution layer 1, maximum pooling layer 1, random inactivation layer 1, convolution layer 2, convolution layer 3, convolution layer 4 and maximum pooling layer 2, and the second convolution network branch includes sequentially connected convolution layer 5, maximum pooling layer 3, random inactivation layer 2, convolution layer 6, convolution layer 7, convolution layer 8 and maximum pooling layer 4, and a batch normalization layer and Le are added after each convolution layer. akyReLU activation function, the features extracted by the two convolutional network branches are fused and then sent to the fully connected layer 1 for flattening after passing through the random dropout layer 3; the time-frequency feature map extraction module includes a convolutional layer 9, a convolutional layer 10, a maximum pooling layer 5 and a random dropout layer 4 connected in sequence, and a batch normalization layer and a LeakyReLU activation function are added after each convolutional layer. The features extracted by the time-frequency feature map extraction module are sent to the fully connected layer 2 for flattening; the fusion module fuses the flattened features of the fully connected layer 1 and the fully connected layer 2 to obtain the multi-view sleep signal features and send them to the classification module; the classification module includes a convolutional layer 11, a convolutional layer 12 and a Softmax layer connected in sequence, and outputs the probability of each sleep stage;

[0013] Step 3.4: training and testing the multi-view hybrid neural network model using the time domain signal in the EEG signal and the time-frequency feature map to obtain a children's sleep stage model;

[0014] Step 4: The server obtains the EEG signal of the target child subject collected by the headband electrode device during sleep, constructs a time-frequency feature map after denoising preprocessing, and inputs the corresponding time domain signal and time-frequency feature map into the child sleep staging model respectively. The model outputs the child sleep staging result, and the Web client displays the child sleep staging result.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] (1) The present invention uses a headband electrode device to collect EEG signals of children during their sleep at night, thereby improving the comfort and cooperation of children during the data collection process, thereby improving the signal quality;

[0017] (2) The present invention introduces a new EMD-based wavelet threshold denoising strategy to denoise the collected children's EEG data, and uses a multi-view feature extraction mechanism to extract signal features from both the time domain and the time-frequency domain. By integrating sleep feature information from different views, the classification accuracy and model performance of the children's sleep staging model are effectively improved, thereby further improving the accuracy of children's sleep staging. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a method for children's sleep staging based on collecting EEG signals with headband electrodes according to an embodiment of the present invention;

[0019] Figure 2 This is the network structure diagram of the multi-view hybrid neural network model. DETAILED DESCRIPTION

[0020] The technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0021] Embodiment 1

[0022] like Figure 1 As shown, the method for children's sleep staging based on collecting EEG signals using headband electrodes proposed in an embodiment of the present invention mainly includes the following steps 1 to 4.

[0023] Step 1: Build the web client and server.

[0024] In this step, the client is built using the Vue framework, Html, CSS and JavaScripts. The server is built using the Flask framework, and the denoising preprocessing module, signal segmentation, time-frequency feature map construction and multi-view hybrid neural network model are deployed on the server.

[0025] Step 2: The headband electrode device collects EEG signals of multiple child subjects during sleep, and sends the collected EEG signals to the server through the Web client.

[0026] Furthermore, the specific process of step 2 is as follows:

[0027] Step 2.1: Ensure that the child subject is in a suitable sleeping environment, comfortable and quiet; ensure that the child's head is clean and free of hair or other objects;

[0028] Step 2.2: According to the design of the acquisition system, the headband electrode device is worn on the head of the child subject to ensure that the electrode is in good contact with the scalp and that the electrode is firmly connected to avoid movement or falling off; the headband electrode device in this embodiment can be implemented by an existing headband device or a headband-type EEG acquisition device, and the headband electrode device mainly includes an elastic band, an electrode array, a Bluetooth communication module, an indicator light and other units;

[0029] Step 2.3: Before starting the acquisition, perform equipment calibration to ensure signal quality and accuracy; perform electrode impedance check to ensure that the connection is good and the impedance is below the set threshold;

[0030] Step 2.4: Start the headband electrode device to record the child's EEG signals throughout the night.

[0031] Step 3: After the server receives the EEG signal sent by the Web client, it executes the following steps 3.1 to 3.4.

[0032] Step 3.1: Use a new wavelet threshold denoising strategy based on EMD to pre-process the EEG signal to obtain sleep EEG data.

[0033] Furthermore, the specific process of performing EMD decomposition on the collected EEG signal (i.e., the original EEG signal) in step 3.1 is as follows:

[0034] Step 3.1.1: Assume the original EEG signal curve is , find the local maximum and local minimum of all original EEG signals;

[0035] Step 3.1.2: For each original EEG signal, use the cubic spline curve to connect the local maximum points of the original EEG signal to form an upper envelope , connect the local minimum points of the original EEG signal to form the lower envelope ;

[0036] Step 3.1.3: Calculate the mean of the upper and lower envelopes , and get the intermediate signal , the calculation formula is as follows:

[0037] ;

[0038] Step 3.1.4: Convert the raw EEG signal With the intermediate signal Subtract and get the remaining signal :

[0039] ;

[0040] Step 3.1.5: Determine the remaining signals Whether the constraints of the intrinsic mode function (IMF) are met. If so, the remaining signal is an IMF component; if it is not satisfied, the remaining signal As the original EEG signal, the remaining signal Re-iterate the above steps 3.1.1 to 3.1.4;

[0041] Step 3.1.6: After obtaining the first IMF component IMF1, use the original EEG signal Subtract the first IMF component IMF1 as a new signal, take the new signal as the original EEG signal, iterate the above steps 3.1.1 to 3.1.5 for the new signal, and obtain the second IMF component IMF2. And so on, complete the EMD decomposition. After the original EEG signal is decomposed by EMD, it is decomposed into several IMF components (high-frequency IMF and low-frequency IMF) and a residual signal;

[0042] Step 3.1.7: Perform wavelet threshold denoising based on particle swarm optimization on the high-frequency IMF components obtained by decomposition;

[0043] Step 3.1.8: Superimpose the denoised high-frequency IMF component with the low-frequency IMF component and the residual signal. The superimposed and reconstructed signal is the denoised child EEG signal. All child EEG signals constitute the final sleep EEG data.

[0044] Step 3.2: Segment and eliminate sleep EEG data and construct a time-frequency feature map.

[0045] Furthermore, the specific process of step 3.2 is as follows:

[0046] Step 3.2.1: Segment the denoised sleep EEG data obtained in step 3.1 into 30-second periods;

[0047] Step 3.2.2: According to the Rechtschaffen & Kales criteria, sleep experts divide the segmented data into periods, and divide the 30-second data into one of the five sleep periods: wakefulness (W), non-rapid eye movement (NREM) sleep stages N1, N2, N3, and rapid eye movement (REM);

[0048] Step 3.2.3: Eliminate some of the divided data, retain only the data from half an hour before falling asleep to half an hour after waking up, and divide the data into training set and test set;

[0049] Step 3.2.4: According to the sampling rate and time resolution of the sleep EEG data, perform a generalized S transform analysis on the sleep EEG data of every 30 seconds, and convert the generalized S transform results of the sleep EEG data of every 30 seconds into a time-frequency feature map by selecting an appropriate window function and frequency and time resolution parameters. For the construction of the time-frequency feature map, the present invention adopts the generalized S transform, and the Gram angular field method can also be used for transformation.

[0050] Step 3.3: Build a multi-view hybrid neural network model.

[0051] The overall structure of the multi-view hybrid neural network model is as follows: Figure 2 As shown in the figure, the multi-view hybrid neural network model uses a two-stream network to extract the features contained in the time domain signal and the time-frequency feature map in the sleep EEG signal, specifically including a time domain double-branch convolution module, a time-frequency feature map extraction module, a fusion module and a classification module.

[0052] The module for extracting features in time domain signals is the time domain dual-branch convolution module. The dual-branch convolution module is mainly composed of two parallel convolution network branches with different kernel sizes. Among them, the convolution layers of different-sized convolution kernels are used to extract different types of features. Small-scale convolution kernels can capture high-frequency features, and large-scale convolution kernels can capture low-frequency features. Specifically, the first convolution network branch includes sequentially connected convolution layer 1, maximum pooling layer 1, random dropout layer 1, convolution layer 2, convolution layer 3, convolution layer 4 and maximum pooling layer 2, and the second convolution network branch includes sequentially connected convolution layer 5, maximum pooling layer 3, random dropout layer 2, convolution layer 6, convolution layer 7, convolution layer 8 and maximum pooling layer 4, and each convolution layer in the two convolution network branches is added with a batch normalization layer and a LeakyReLU activation function. The features extracted by the two convolution network branches are fused and then sent to the fully connected layer 1 through the random dropout layer 3 for flattening.

[0053] The module for extracting features in the time-frequency feature map is the time-frequency feature map extraction module, which includes convolution layers 9 and 10 for feature learning, a maximum pooling layer 5 and a random dropout layer 4 for feature dimensionality reduction, and a batch normalization layer and a LeakyReLU activation function are added after each convolution layer in the time-frequency feature map extraction module. The features extracted by the time-frequency feature map extraction module are sent to the fully connected layer 2 for flattening.

[0054] The fusion module fuses the flattened features of the fully connected layer 1 and the fully connected layer 2 to obtain the multi-view sleep signal features.

[0055] The classification module includes a convolution layer 11, a convolution layer 12 and a Softmax layer connected in sequence. The obtained multi-view sleep signal features are input into the classification module, and the probability of each sleep stage is output, thereby obtaining the staging results of children's sleep.

[0056] In the specific operation, the input size of the time-domain dual-branch convolution module for time-domain signal feature extraction is (7500, 3). Each convolution network branch in the time-domain dual-branch convolution module contains four convolution layers, two pooling layers and a random dropout layer. The first convolution network branch is a small core branch, and the second convolution network branch is a large core branch. The size of convolution layer 1 in the small core branch is (64, 50, 8), (64, 50, 8) means that the convolution kernel size is 50, the number of convolution kernels is 64, and the step size is 8. The sizes of convolution layers 2 and 3 in the small core branch are (128, 8, 1), and the size of convolution layer 4 is (64, 8, 1). The size of the maximum pooling layer 1 in the small core branch is (8, 8), (8, 8) means that the convolution kernel size is 8, the step size is 8, and the size of the maximum pooling layer 2 is (4, 4). The size of convolution layer 5 in the large core branch is (64, 500, 64), the size of convolution layer 6 and convolution layer 7 is (128, 6, 1), the size of convolution layer 8 is (64, 6, 1), the size of maximum pooling layer 3 is (4, 4), and the size of maximum pooling layer 4 is (2, 2). The inactivation probability of random dropout layer 1 and random dropout layer 2 in the small core branch and the large core branch is set to 0.5. Batch normalization layer and LeakyReLU activation function are added after each convolution layer in the small core branch and the large core branch.

[0057] The input size of the time-frequency feature map extraction module for feature extraction of time-frequency feature maps is (30, 125, 3), where the size of convolution layer 9 is (64, 3, 1), the size of convolution layer 10 is (64, 5, 1), the activation function is LeakyReLU, the size and stride of the maximum pooling layer 5 are 5, and the inactivation probability of the random inactivation layer 4 is 0.5.

[0058] Step 3.4: Use the time domain signals and time-frequency feature maps in the EEG signals to train and test the multi-view hybrid neural network model to obtain the children's sleep staging model and realize automatic sleep staging based on children's EEG signals.

[0059] Step 3.4.1: Training phase.

[0060] Step 3.4.1.1: When training the multi-view hybrid neural network model, use the sleep EEG data in the training set of child subjects 1 to child subjects N-1 to construct the time domain signal and time-frequency feature map respectively;

[0061] Step 3.4.1.2: Then, the time domain signal and the time-frequency feature map are respectively input into the multi-view hybrid neural network model, that is, the time domain signal is input into the time domain double-branch convolution module, and the time-frequency feature map is input into the time-frequency feature map extraction module. The obtained result is flattened through the fully connected layer, and the flattened result is fused through the fusion module to obtain the multi-view sleep signal feature;

[0062] Step 3.4.1.3: Input the multi-view sleep signal features into the classification module, output the sleep staging results through the classification module, input the sleep staging results into the cross entropy loss function to obtain the corresponding loss function, use the stochastic gradient descent method to optimize the loss function and adjust the model parameters, and finally obtain the trained children's sleep staging model.

[0063] Step 3.4.2: Testing phase.

[0064] Step 3.4.2.1: In the specific operation, the sleep EEG data in the test set of child subject N is used to construct the time domain signal and time-frequency feature map respectively;

[0065] Step 3.4.2.2: The constructed time domain signal and time-frequency feature map are then input into the trained multi-view hybrid neural network model, namely the children's sleep staging model, and the children's sleep staging model outputs the final children's sleep staging result.

[0066] Step 4: When using the child sleep staging model to stage the child's sleep, the headband electrode device collects the EEG signals of the target child subject during sleep and sends them to the server through the Web client. After the server obtains the EEG signals of the target child subject collected by the headband electrode device during sleep, it executes steps 3.1-3.2. The EEG signals are pre-processed by denoising to construct the corresponding time-frequency feature graph, and the child sleep staging model is called to automatically stage the child's sleep EEG data. The time domain signal and the time-frequency feature graph are input into the child sleep staging model. The child sleep staging model outputs the child sleep staging results, and the Web client displays the child sleep staging results. The visualization results of the Web client also include the whole night EEG of the child's sleep data.

[0067] The present invention uses a children's sleep EEG data acquisition method based on headband electrodes to collect children's whole-night sleep data, uses a new wavelet threshold denoising strategy based on EMD to denoise the collected children's EEG data, and uses a multi-view hybrid neural network to perform children's sleep stages on the denoised EEG signals. While improving the comfort and cooperation of children in the acquisition or monitoring process, thereby improving signal quality, the multi-view feature extraction mechanism is used to extract signal features from two perspectives, namely, the time domain and the time-frequency domain. By integrating sleep feature information from different views, the classification accuracy and model performance of the children's sleep staging model are effectively improved, thereby further improving the accuracy of children's sleep staging.

[0068] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0069] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for children's sleep staging based on collecting EEG signals using headband electrodes, characterized in that: The following steps are involved: Step 1: Build the web client and server; Step 2: The headband electrode device collects EEG signals of multiple child subjects during sleep, and sends the EEG signals to the server through the Web client; Step 3: The server performs the following steps: Step 3.1: Using a new wavelet threshold denoising strategy based on EMD to pre-process the EEG signal to obtain sleep EEG data; Step 3.2: segmenting and eliminating the sleep EEG data to construct a time-frequency feature map; Step 3.3: Construct a multi-view hybrid neural network model, which includes a time-domain dual-branch convolution module, a time-frequency feature map extraction module, a fusion module and a classification module. The time-domain dual-branch convolution module takes the time domain signal as input, and includes two parallel convolution network branches with different kernel sizes. The first convolution network branch includes sequentially connected convolution layer 1, maximum pooling layer 1, random inactivation layer 1, convolution layer 2, convolution layer 3, convolution layer 4 and maximum pooling layer 2, and the second convolution network branch includes sequentially connected convolution layer 5, maximum pooling layer 3, random inactivation layer 2, convolution layer 6, convolution layer 7, convolution layer 8 and maximum pooling layer 4, and a batch normalization layer and Le are added after each convolution layer. akyReLU activation function, the features extracted by the two convolutional network branches are fused and then sent to the fully connected layer 1 for flattening after passing through the random dropout layer 3; the time-frequency feature map extraction module includes a convolutional layer 9, a convolutional layer 10, a maximum pooling layer 5 and a random dropout layer 4 connected in sequence, and a batch normalization layer and a LeakyReLU activation function are added after each convolutional layer. The features extracted by the time-frequency feature map extraction module are sent to the fully connected layer 2 for flattening; the fusion module fuses the flattened features of the fully connected layer 1 and the fully connected layer 2 to obtain the multi-view sleep signal features and send them to the classification module; the classification module includes a convolutional layer 11, a convolutional layer 12 and a Softmax layer connected in sequence, and outputs the probability of each sleep stage; Step 3.4: training and testing the multi-view hybrid neural network model using the time domain signal in the EEG signal and the time-frequency feature map to obtain a children's sleep stage model; Step 4: The server obtains the EEG signal of the target child subject collected by the headband electrode device during sleep, constructs a time-frequency feature map after denoising preprocessing, and inputs the corresponding time domain signal and time-frequency feature map into the child sleep staging model respectively. The model outputs the child sleep staging result, and the Web client displays the child sleep staging result.

2. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1, characterized in that: In step 2, before the headband electrode device starts collecting EEG signals, device calibration and electrode impedance check are performed.

3. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1 or 2, characterized in that: Step 3.1 includes the following steps: Step 3.1.1: Find the local maximum and local minimum of all EEG signals; Step 3.1.2: For each original EEG signal, use a cubic spline curve to connect the local maximum points of the EEG signal to form an upper envelope , connect the local minimum points of the EEG signal to form the lower envelope ; Step 3.1.3: Calculate the mean of the upper and lower envelopes , and get the intermediate signal ; Step 3.1.4: Subtract the EEG signal from the intermediate signal to obtain the remaining signal; Step 3.1.5: Determine whether the remaining signal satisfies the constraints of the intrinsic mode function. If so, the remaining signal is an IMF component; otherwise, reiterate the above steps 3.1.1 to 3.1.4 for the remaining signal; Step 3.1.6: After obtaining the first IMF component, subtract the first IMF component from the EEG signal, and use the obtained signal as the new signal. Re-iterate the above steps 3.1.1 to 3.1.5 for the new signal to obtain the second IMF component, and so on, to complete the EMD decomposition and decompose the EEG signal into several IMF components and a residual signal; Step 3.1.7: Perform wavelet threshold denoising based on particle swarm optimization on the high-frequency IMF components obtained by decomposition; Step 3.1.8: Superimpose the denoised high-frequency IMF component, the low-frequency IMF component and the residual signal, and the superimposed and reconstructed signal is the obtained sleep EEG data.

4. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1 or 2, characterized in that: Step 3.2 includes the following steps: Step 3.2.1: segmenting the sleep EEG data into 30-second periods; Step 3.2.2: According to the Rechtschaffen & Kales standard, sleep experts divide the segmented data into five sleep periods: wakefulness, non-rapid eye movement sleep stages N1, N2, N3, and rapid eye movement sleep; Step 3.2.3: Eliminate some of the divided data, retain only the data from half an hour before falling asleep to half an hour after waking up, and divide the data into training set and test set; Step 3.2.4: According to the sampling rate and time resolution of the sleep EEG data, perform generalized S transform analysis on the sleep EEG data of every 30 seconds, and convert each generalized S transform result into the corresponding time-frequency feature map.

5. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 4, characterized in that: The process of training the multi-view hybrid neural network model includes the following steps: Step 3.4.1.1: Use the sleep EEG data in the training set of child subjects 1 to child subjects N-1 to construct the time domain signal and time-frequency feature map respectively; Step 3.4.1.2: Input the time domain signal into the time domain dual-branch convolution module, input the time-frequency feature map into the time-frequency feature map extraction module, flatten the obtained results through the fully connected layer and fuse them through the fusion module to obtain the multi-view sleep signal features; Step 3.4.1.3: Input the multi-view sleep signal features into the classification module, output the sleep staging results through the classification module, input the sleep staging results into the cross entropy loss function to obtain the corresponding loss function, use the stochastic gradient descent method to optimize the loss function and adjust the model parameters, and finally obtain the trained children's sleep staging model.

6. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 5, characterized in that: The process of testing the multi-view hybrid neural network model includes the following steps: Step 3.4.2.1: Use the sleep EEG data in the test set of child subject N to construct the time domain signal and time-frequency feature map respectively; Step 3.4.2.2: Input the constructed time domain signal and time-frequency feature map into the children's sleep staging model, and the model outputs the final children's sleep staging results.

7. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1 or 2, characterized in that: The input size of each convolutional network branch is (7500, 3); The first convolutional network branch is a small core branch. The sizes of convolutional layer 1, convolutional layer 2, convolutional layer 3, and convolutional layer 4 in the small core branch are (64, 50, 8), (128, 8, 1), (128, 8, 1), and (64, 8, 1), respectively. The sizes of maximum pooling layer 1 and maximum pooling layer 2 are (8, 8) and (4, 4), respectively. The second convolutional network branch is the large core branch. The sizes of convolutional layers 5, 6, 7, and 8 in the large core branch are (64, 500, 64), (128, 6, 1), (128, 6, 1), and (64, 6, 1), respectively. The sizes of maximum pooling layers 3 and 4 are (4, 4) and (2, 2), respectively. The dropout probability of dropout layer 1 and dropout layer 2 is 0.

5.

8. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1 or 2, characterized in that: The input size of the time-frequency feature map extraction module is (30, 125, 3), the sizes of convolutional layer 9 and convolutional layer 10 are (64, 3, 1) and (64, 5, 1) respectively, the size of the maximum pooling layer 5 is (5, 5), and the inactivation probability of the random inactivation layer 4 is 0.

5.

9. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1 or 2, characterized in that: Using the Vue framework, build the web client using Html, CSS, and JavaScripts.

10. The method for children's sleep staging based on collecting EEG signals using headband electrodes according to claim 1 or 2, characterized in that: The server side is built using the Flask framework.

Images