Automatic sleep staging method and system based on multi-scale feature extraction

Abstract

The invention discloses an automatic sleep staging method and system based on multi-scale feature extraction. The method comprises the following steps that an electroencephalogram time-frequency diagram of a testee is obtained and preprocessed; building a multi-scale automatic sleep staging model based on an Inception network and an SE module; and inputting the preprocessed electroencephalogram time-frequency graph into the multi-scale automatic sleep staging model to obtain a sleep staging result. According to the method, through a deep feature extraction model, the overall channel features can be enhanced by combining multi-scale convolution with an SE channel attention module, the model performance is enhanced, and the sleep staging accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of automatic sleep staging, and in particular relates to an automatic sleep staging method and system based on multi-scale feature extraction. Background Art

[0002] In the field of sleep state research, sleep staging has always been a hot topic in healthy sleep research. Although feature selection and data enhancement can effectively improve the accuracy of sleep staging, a single convolutional neural network (CNN) cannot extract deeper image features. Currently, most of the methods people use to improve deep learning networks are to simply and directly increase the number of layers and width, thereby improving the network's feature extraction capabilities. However, this approach will lead to the problem of gradient vanishing easily during back propagation as the number of network layers increases and the number of parameters increases, making network training more difficult. Therefore, in order to improve the accuracy of sleep staging, it is necessary to further optimize network training and extract deeper image features. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention proposes an automatic sleep staging method and system based on multi-scale feature extraction. Through a deep feature extraction model, multi-scale convolution combined with the SE channel attention module can be used to strengthen the overall channel characteristics, enhance model performance, and improve the accuracy of sleep staging.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] An automatic sleep staging method based on multi-scale feature extraction comprises the following steps:

[0006] Obtain the EEG frequency graph of the subject and perform preprocessing;

[0007] Build a multi-scale automatic sleep staging model based on the Inception network and SE module;

[0008] The preprocessed EEG frequency graph is input into the multi-scale automatic sleep staging model to obtain a sleep staging result.

[0009] Preferably, the multi-scale automatic sleep staging model includes: a feature cascade group module and a classification module;

[0010] The feature cascade group module is used to extract features of different dimensions using convolution blocks of different scales, and obtain a final feature map by splicing features of different dimensions;

[0011] The classification module is used to obtain a sleep staging result based on the final feature map.

[0012] Preferably, the feature cascade group module comprises three parts, namely: a shallow feature extraction submodule, a multi-scale feature extraction submodule, and a deep feature extraction submodule;

[0013] The shallow feature extraction submodule is used to extract shallow features and perform dimensionality reduction based on the preprocessed EEG time-frequency graph to obtain a shallow feature graph;

[0014] The multi-scale feature extraction submodule is used to extract image features of different scales based on the shallow feature map and fuse them to obtain a fused feature map;

[0015] The deep feature extraction submodule is used to obtain a final feature map based on the fused feature map after three convolutions and feature fusion.

[0016] Preferably, based on the preprocessed EEG time-frequency graph, shallow features are extracted and dimension reduction is performed to obtain a shallow feature graph, and the process includes:

[0017] The preprocessed EEG time-frequency map is transformed into a two-dimensional time-frequency map of 96*96*3 size through wavelet time-frequency transformation as input, and shallow feature extraction is performed through a two-dimensional convolution layer composed of two convolution kernels of size 3*3*64 connected in series. The feature dimension reduction is performed using the maximum pooling layer, and the output scale is a feature map of 48*48*64.

[0018] Preferably, based on the shallow feature map, image features of different scales are extracted and fused to obtain the fused feature map, and the process includes:

[0019] The feature map with a scale of 48*48*64 output by the shallow feature extraction module is taken as input, and image features of different scales are extracted through convolution kernels of different sizes: 3*3, 5*5, and 7*7. Through parallel operation and 1*1 convolution kernel, the fused feature map after multi-scale feature extraction is obtained.

[0020] Preferably, based on the fused feature map, after three convolutions and feature fusion, the process of obtaining the final feature map includes:

[0021] Combined with the VGG network, a 3*3 convolution kernel is used to build a deep network; and based on the deep network, the SE attention mechanism is combined to dynamically learn the importance of each channel, assign different weight values ​​to each channel, and obtain the final feature map.

[0022] The present invention also provides an automatic sleep staging system based on multi-scale feature extraction, comprising: an acquisition module, a construction module and a classification module;

[0023] The acquisition module is used to obtain the EEG time-frequency diagram of the subject and perform preprocessing;

[0024] The construction module is used to build a multi-scale automatic sleep staging model based on the Inception network and the SE module;

[0025] The classification module is used to input the preprocessed EEG frequency graph into the multi-scale automatic sleep staging model to obtain a sleep staging result.

[0026] Preferably, the multi-scale automatic sleep staging model includes: a feature cascade group module and a classification module;

[0027] The feature cascade group module is used to extract features of different dimensions using convolution blocks of different scales, and obtain a final feature map by splicing features of different dimensions;

[0028] The classification module is used to obtain a sleep staging result based on the final feature map.

[0029] Preferably, the feature cascade group module comprises three parts, namely: a shallow feature extraction submodule, a multi-scale feature extraction submodule, and a deep feature extraction submodule;

[0030] The shallow feature extraction submodule is used to extract shallow features and perform dimensionality reduction based on the preprocessed EEG time-frequency graph to obtain a shallow feature graph;

[0031] The multi-scale feature extraction submodule is used to extract image features of different scales based on the shallow feature map and fuse them to obtain a fused feature map;

[0032] The deep feature extraction submodule is used to obtain a final feature map based on the fused feature map after three convolutions and feature fusion.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention provides an automatic sleep staging method and system based on multi-scale feature extraction. Based on the Inception series model, the features of different scales of the image can be extracted according to convolution kernels of different sizes and pooling operations, and then these features can be merged together to enrich the network's understanding of the image; the SE module can be used to introduce the attention mechanism between channels, adaptively learn and strengthen the attention to important features, and suppress the attention to unimportant features; a multi-scale feature extraction staging model can be built by combining the Inception multi-scale feature extraction module, the SE module and the VGG network structure on the basis of the original CNN network, and the effectiveness of the improved multi-scale feature extraction model is verified by comparing the convolution kernel combinations of different sizes, different model combinations, the classification effects of different network models, and the classification accuracy before and after the data set is averaged. This application improves the accuracy of automatic sleep staging by combining and optimizing the original network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 A schematic flow chart of an automatic sleep staging method based on multi-scale feature extraction according to an embodiment of the present invention;

[0037] Figure 2 Schematic diagram of a shallow feature extraction module according to an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of a multi-scale feature extraction module according to an embodiment of the present invention;

[0039] Figure 4 Schematic diagram of a deep feature extraction module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Embodiment 1

[0043] like Figure 1 As shown, the present invention provides an automatic sleep staging method based on multi-scale feature extraction, based on the Inception series model, enriching the network's understanding of images; based on the SE attention module, introducing the attention mechanism between channels, learning the importance weight of each channel to dynamically adjust the feature representation; combining the two parts, by comparing the feature extraction capabilities of the models under different convolution kernel sizes and number combinations, and comparing the feature extraction capabilities of the models under different module combinations to find the best network model, and by comparing the classification results of multiple network models to verify the effect of the improved network. The specific implementation process is as follows:

[0044] Obtain the EEG frequency graph of the subject and perform preprocessing;

[0045] Build a multi-scale automatic sleep staging model based on the Inception network and SE module;

[0046] The preprocessed EEG frequency graph is input into the multi-scale automatic sleep staging model to obtain a sleep staging result.

[0047] In this embodiment, a multi-scale automatic sleep staging model is built based on the Inception network and the SE module. The 96*96*3 EEG frequency map is used as input, and it is divided into 5 different sleep stages through a series of feature extraction modules. After continuous feature extraction, a more effective feature map is obtained, which further improves the accuracy of sleep staging.

[0048] In this embodiment, the multi-scale automatic sleep staging model includes: a feature cascade group module and a classification module;

[0049] The feature cascade group module is used to extract features of different dimensions using convolution blocks of different scales, and obtain the final feature map by splicing features of different dimensions;

[0050] The classification module is used to obtain the result of sleep staging based on the final feature map.

[0051] In this embodiment, the feature cascade group module includes three parts, namely: a shallow feature extraction submodule, a multi-scale feature extraction submodule, and a deep feature extraction submodule;

[0052] The shallow feature extraction submodule is used to extract shallow features and perform dimensionality reduction based on the preprocessed EEG time-frequency graph to obtain a shallow feature graph;

[0053] The multi-scale feature extraction submodule is used to extract image features of different scales based on the shallow feature map and fuse them to obtain a fused feature map;

[0054] The deep feature extraction submodule is used to obtain the final feature map after three convolutions and feature fusion based on the fused feature map.

[0055] Specifically, the present invention is based on the Sleep-EDF Database Expanded database. The present invention uses 40 people's data from the Fpz-Cz lead in the healthy population (SC) data set for experimentation. In 2007, the American Academy of Sleep Medicine (AASM) formulated a new sleep staging standard, and the sleep state was divided into wakefulness W, REM (rapid eye movement), NREM (non-rapid eye movement), among which NREM includes N1, N2, and N3. The AASM summarizes the characteristics of EEG signals in different sleep stages as follows:

[0056] Wake period: The awake period is dominated by a mixed wave voltage of 10 to 30 μv, with a frequency range of 16 to 25 Hz.

[0057] Stage N1: The stage of falling asleep, which accounts for about 5% of the total sleep time of the whole night and lasts for a shorter time during normal sleep.

[0058] Stage N2: Light sleep stage, which indicates that the human body has truly entered sleep. The light sleep stage accounts for about 44% to 55% of the total sleep time.

[0059] N3 stage: deep sleep. During this period, sleep spindles gradually disappear in the sleep EEG, high-amplitude slow waves account for 20% to 50%, and spindles may continue to appear.

[0060] REM stage: It is the main period of dreaming, characterized by low amplitude and mixed frequency. Obvious sawtooth waves and rapid eye movements can be observed.

[0061] The N3 and N4 periods were merged into the N3 stage data, and the data of irrelevant stages were deleted. In order to avoid the collection of a large amount of useless data due to the long wake period of some subjects, the sleep electroencephalogram (EEG) signal from 15 minutes before sleep to 15 minutes after waking up was uniformly intercepted before processing the data. The eighth-order Butterworth low-pass filter and wavelet denoising method were used for preprocessing.

[0062] Since continuous wavelet transform can provide local characteristic information of the signal in time and frequency, for non-stationary time series signals such as sleep EEG signals, continuous wavelet transform can well capture the instantaneous dynamic changes of signal frequency. The present invention uses continuous wavelet transform to perform time-frequency transformation on the pre-processed EEG signal.

[0063] Convolutional neural network (CNN) is a neural network designed specifically for processing image data. It can capture local features by using convolutional layers and reduce the dimension of data by using pooling layers. CNN network has the characteristics of local perception and sparse interaction, parameter sharing, automatic feature learning, spatial invariance, model compression, hierarchical feature extraction, etc. In this method, it is proposed to build an automatic sleep staging model based on CWT-CNN based on the advantages of CNN network in image data classification.

[0064] Simple CNN networks have limitations in extracting deep features of images. Increasing the number of network layers blindly can easily lead to problems such as gradient vanishing. A multi-scale automatic sleep staging model is built based on the Inception network and SE module. Convolutional blocks of different scales are used to extract features of different dimensions, and the splicing of features brings richer feature information.

[0065] The multi-scale feature extraction network mainly consists of two parts: a feature cascade group module and a classification module. The feature cascade group module consists of three parts: shallow feature extraction, multi-scale feature extraction, and deep feature extraction.

[0066] The two-dimensional time-frequency graph obtained after preprocessing and time-frequency analysis is used as the input of the feature cascade group. After passing through the cascade feature module, the fully connected layer of the classification module is used to output the sleep stage results. The feature cascade group module and the classification module together form a multi-scale feature network.

[0067] (1) Shallow feature extraction module: The preprocessed EEG signal is transformed into a 96*96*3 two-dimensional time-frequency map as input through wavelet time-frequency transformation, and shallow feature extraction is performed through a two-dimensional convolution layer composed of two convolution kernels of size 3*3*64 connected in series. The feature dimension is reduced using the maximum pooling layer, and the output feature map scale is 48*48*64. Figure 2 shown.

[0068] (2) Multi-scale feature extraction module: The feature map of size 48*48*64 output by the shallow feature extraction module is used as input. In order to capture the features of different scales in the image, the image features of different scales are extracted through convolution kernels of different sizes such as 3*3, 5*5, 7*7, etc. Through parallel operation and the use of 1*1 convolution kernel, the depth and width of the network can be increased while maintaining high efficiency, which helps to reduce the number of parameters and computational complexity of the model, making the model more lightweight. Figure 3 shown.

[0069] (3) Deep feature extraction module: Combined with the VGG network, a smaller 3*3 convolution kernel is used. By stacking multiple small convolution kernels to replace a large convolution kernel, it helps to reduce the number of parameters and the amount of calculation, and build a deeper network. In addition, the SE attention mechanism is combined with the deep network to dynamically learn the importance of each channel and assign different weight values ​​to each channel to enhance the model's attention to important features, thereby improving the model's performance and classification accuracy.

[0070] This module takes the fused feature map after multi-scale feature extraction as input, and after three convolutions and feature fusion, the final feature map (see Figure 2) is obtained. Figure 1 As shown in Figure 2, the classification module is sent to obtain the final classification result. Figure 4 shown.

[0071] The fully connected layer FC of the classification module is usually used in the neural network to comprehensively process the learned features and make the final classification decision. Before the fully connected layer, the present invention has adopted multi-scale to extract deep features. The feature map is flattened into a one-dimensional vector to connect each neuron in the fully connected layer to all activation values ​​of the previous layer. Each connection has a weight, and the output of the neuron is the sum of these weighted sums, plus a bias term. The specific formula is as follows:

[0072] Output=W×Input+b (1)

[0073] Among them, W is the weight matrix, InputInput is the input feature, and b is the bias term.

[0074] An activation function such as ReLU (Rectified Linear Unit) or Sigmoid is usually applied to the output of the fully connected layer. ReLU is a commonly used nonlinear activation function that can increase the nonlinear ability of the model and help the network learn more complex patterns. In multi-classification problems, the Softmax function is used. The Softmax function converts the value of the output layer into a probability distribution, and the sum of the probabilities of each category is 1.

[0075] Appendix Table 1 lists the detailed feature map dimensional change process of the present invention, and clearly describes the overall framework of the model. The automatic sleep staging model with multi-scale feature extraction takes a 96*96*3 EEG time-frequency map as input, and divides it into 5 different sleep stages through a series of feature extraction modules. After continuous feature extraction, a more effective feature map is obtained, which further improves the accuracy of sleep staging.

[0076] Table 1 Multi-scale classification model parameters

[0077]

[0078] Embodiment 2

[0079] The present invention also provides an automatic sleep staging system based on multi-scale feature extraction, comprising: an acquisition module, a construction module and a classification module;

[0080] The acquisition module is used to obtain the EEG time-frequency diagram of the subject and perform preprocessing;

[0081] The construction module is used to build a multi-scale automatic sleep staging model based on the Inception network and SE module;

[0082] The classification module is used to input the preprocessed EEG time-frequency map into the multi-scale automatic sleep staging model to obtain a sleep staging result.

[0083] In this embodiment, the multi-scale automatic sleep staging model includes: a feature cascade group module and a classification module;

[0084] The feature cascade group module is used to extract features of different dimensions using convolution blocks of different scales, and obtain the final feature map by splicing features of different dimensions;

[0085] The classification module is used to obtain the result of sleep staging based on the final feature map.

[0086] In this embodiment, the feature cascade group module includes three parts, namely: a shallow feature extraction submodule, a multi-scale feature extraction submodule, and a deep feature extraction submodule;

[0087] The shallow feature extraction submodule is used to extract shallow features and perform dimensionality reduction based on the preprocessed EEG time-frequency graph to obtain a shallow feature graph;

[0088] The multi-scale feature extraction submodule is used to extract image features of different scales based on the shallow feature map and fuse them to obtain a fused feature map;

[0089] The deep feature extraction submodule is used to obtain the final feature map after three convolutions and feature fusion based on the fused feature map.

[0090] In this embodiment, based on the preprocessed EEG time-frequency graph, shallow features are extracted and dimension reduction is performed to obtain a shallow feature graph, and the process includes:

[0091] The preprocessed EEG time-frequency map is transformed into a two-dimensional time-frequency map of 96*96*3 size through wavelet time-frequency transformation as input, and shallow feature extraction is performed through a two-dimensional convolution layer composed of two convolution kernels of size 3*3*64 connected in series. The feature dimension reduction is performed using the maximum pooling layer, and the output scale is a feature map of 48*48*64.

[0092] In this embodiment, based on the shallow feature map, image features of different scales are extracted and fused to obtain the fused feature map, which includes:

[0093] The feature map with a scale of 48*48*64 output by the shallow feature extraction module is taken as input, and image features of different scales are extracted through convolution kernels of different sizes: 3*3, 5*5, and 7*7. Through parallel operation and 1*1 convolution kernel, the fused feature map after multi-scale feature extraction is obtained.

[0094] In this embodiment, based on the fused feature map, after three convolutions and feature fusion, the process of obtaining the final feature map includes:

[0095] Combined with the VGG network, a 3*3 convolution kernel is used to build a deep network; and based on the deep network, the SE attention mechanism is combined to dynamically learn the importance of each channel, assign different weight values ​​to each channel, and obtain the final feature map.

[0096] Embodiment 3

[0097] Based on the multi-scale automatic sleep staging network constructed in Example 1, the optimal network model is found by comparing the feature extraction capabilities of the models under different combinations of convolution kernel sizes and numbers, and the feature extraction capabilities of the models under different module combinations. At the same time, the effect of the improved network is verified by comparing the classification results of multiple network models.

[0098] (1) Comparison of model feature extraction capabilities under different convolution kernel sizes and number combinations

[0099] Based on the Inception multi-scale feature extraction network model, a multi-scale convolution kernel parallel structure is used to extract image features of different scales. The analysis shows that the classification accuracy has been significantly improved compared with the traditional CNN network. The convolution kernel combination with the best classification effect has an average classification accuracy of 72.4%. Among them, the classification accuracy of W, N2, N3, and REM has reached more than 70%.

[0100] (2) Comparison of the impact of multi-convolution and SE modules

[0101] In order to explore the influence of SE attention module on classification model, classification experiments were conducted on four models, namely original CNN network classification model, CNN network model + multi-convolution Inception model, CNN network model + SE attention module, CNN + multi-convolution Inception model + SE attention module, to find the model combination with the best classification effect. After analysis, the fourth model combination achieved the highest average classification accuracy; the combination with multi-convolution Inception model achieved better classification effect than the original CNN network model; and in the classification results of CNN model combined with SE module, only N3 deep sleep stage achieved good staging accuracy, and the staging accuracy of other stages even decreased to a certain extent. The classification model with multi-convolution Inception model and SE attention module added at the same time made up for the deficiency of adding SE attention module alone, and solved the problem of poor staging effect of some stages caused by adding SE module alone.

[0102] (3) Comparison of results of different classification models

[0103] The VGG network is one of the most popular CNN models. It uses smaller convolution kernels (usually 3*3) and stacks multiple small convolution kernels to improve the nonlinear representation ability of the network while reducing the number of parameters and the amount of calculation. This proposal uses VGG16 as the staging network model.

[0104] To solve the problem of gradient disappearance and network degradation during deep network training, ResNet introduces residual connections to achieve cross-layer information transfer, so that good performance and training effects can be maintained in deeper networks. This proposal uses ResNet34 as the sleep staging network model.

[0105] The classification accuracy of different network models was compared to verify the improvement effect of the improved network staging model compared with the traditional mainstream neural network. After analysis, the classification accuracy of the improved multi-scale feature extraction sleep staging model was 74.5% for stage W, 45.8% for stage N1, 77.2% for stage N2, 81.9% for stage N3, and 74.7% for stage N4, which were all higher than the other three feature extraction networks, verifying the effectiveness of the improvement of the multi-scale feature extraction network.

[0106] (4) Comparison of classification results before and after data set averaging

[0107] To further verify the effectiveness of the improved network with multi-scale feature extraction and improve the accuracy of sleep staging, sleep staging was performed using the sleep datasets before and after dataset averaging as inputs, respectively, using the CNN networks before and after the improvement. After analysis, the accuracy of sleep staging at each stage obtained using the improved CNN network was higher than the classification accuracy of the original CNN network. Compared with the 70.9% staging accuracy obtained by using the original CNN network for staging before dataset averaging, the improved multi-scale feature extraction staging model obtained an accuracy of 77.8% after data averaging, an increase of 6.9%, further verifying the feasibility of dataset averaging and improved multi-scale feature extraction models.

[0108] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An automatic sleep staging method based on multi-scale feature extraction, characterized in that: The following steps are involved: Obtain the EEG frequency graph of the subject and perform preprocessing; Build a multi-scale automatic sleep staging model based on the Inception network and SE module; The preprocessed EEG frequency graph is input into the multi-scale automatic sleep staging model to obtain a sleep staging result.

2. The automatic sleep staging method based on multi-scale feature extraction according to claim 1, characterized in that: The multi-scale automatic sleep staging model includes: a feature cascade group module and a classification module; The feature cascade group module is used to extract features of different dimensions using convolution blocks of different scales, and obtain a final feature map by splicing features of different dimensions; The classification module is used to obtain a sleep staging result based on the final feature map.

3. The automatic sleep staging method based on multi-scale feature extraction according to claim 1, characterized in that: The feature cascade group module includes three parts, namely: a shallow feature extraction submodule, a multi-scale feature extraction submodule, and a deep feature extraction submodule; The shallow feature extraction submodule is used to extract shallow features and perform dimensionality reduction based on the preprocessed EEG time-frequency graph to obtain a shallow feature graph; The multi-scale feature extraction submodule is used to extract image features of different scales based on the shallow feature map and fuse them to obtain a fused feature map; The deep feature extraction submodule is used to obtain a final feature map based on the fused feature map after three convolutions and feature fusion.

4. The automatic sleep staging method based on multi-scale feature extraction according to claim 3 is characterized in that: Based on the preprocessed EEG time-frequency graph, shallow features are extracted and dimension reduction is performed. The process of obtaining shallow feature graph includes: The preprocessed EEG time-frequency map is transformed into a two-dimensional time-frequency map of 96*96*3 size through wavelet time-frequency transformation as input, and shallow feature extraction is performed through a two-dimensional convolution layer composed of two convolution kernels of size 3*3*64 connected in series. The feature dimension reduction is performed using the maximum pooling layer, and the output scale is a feature map of 48*48*64.

5. The automatic sleep staging method based on multi-scale feature extraction according to claim 3, characterized in that: Based on the shallow feature map, the image features of different scales are extracted and fused to obtain the fused feature map, and the process includes: The feature map with a scale of 48*48*64 output by the shallow feature extraction module is taken as input, and image features of different scales are extracted through convolution kernels of different sizes: 3*3, 5*5, and 7*7. Through parallel operation and 1*1 convolution kernel, the fused feature map after multi-scale feature extraction is obtained.

6. The automatic sleep staging method based on multi-scale feature extraction according to claim 3, characterized in that: Based on the fused feature map, after three convolutions and feature fusion, the process of obtaining the final feature map includes: Combined with the VGG network, a 3*3 convolution kernel is used to build a deep network; and based on the deep network, the SE attention mechanism is combined to dynamically learn the importance of each channel, assign different weight values ​​to each channel, and obtain the final feature map.

7. An automatic sleep staging system based on multi-scale feature extraction, characterized in that: include: Acquisition module, construction module and classification module; The acquisition module is used to obtain the EEG time-frequency diagram of the subject and perform preprocessing; The construction module is used to build a multi-scale automatic sleep staging model based on the Inception network and the SE module; The classification module is used to input the preprocessed EEG frequency graph into the multi-scale automatic sleep staging model to obtain a sleep staging result.

8. The automatic sleep staging system based on multi-scale feature extraction according to claim 7, characterized in that: The multi-scale automatic sleep staging model includes: a feature cascade group module and a classification module; The feature cascade group module is used to extract features of different dimensions using convolution blocks of different scales, and obtain a final feature map by splicing features of different dimensions; The classification module is used to obtain a sleep staging result based on the final feature map.

9. The automatic sleep staging system based on multi-scale feature extraction according to claim 7, characterized in that: The feature cascade group module includes three parts, namely: a shallow feature extraction submodule, a multi-scale feature extraction submodule, and a deep feature extraction submodule; The shallow feature extraction submodule is used to extract shallow features and perform dimensionality reduction based on the preprocessed EEG time-frequency graph to obtain a shallow feature graph; The multi-scale feature extraction submodule is used to extract image features of different scales based on the shallow feature map and fuse them to obtain a fused feature map; The deep feature extraction submodule is used to obtain a final feature map based on the fused feature map after three convolutions and feature fusion.

Images