A Single-Channel EEG Sleep Classification Model and Its Construction Method

Abstract

The application provides a single-channel electroencephalogram sleep classification model construction method, comprising the following steps: S1, obtaining original PSG data, wherein the PSG data contains multiple records, each record contains multiple channel data, and each record corresponds to a sleep stage classification label; S2, taking a preset electroencephalogram channel as a target single channel, taking multiple other channels as associated channels, and screening target single channel data and original associated channel data from each record in the original PSG data to form a training data set; S3, constructing an initial classification model, wherein the initial model comprises a feature extraction module, a generative adversarial unit, a classification unit and a fusion module; and S4, training the initial classification model to convergence to obtain a model comprising a feature extraction module, multiple generators, a classification unit and a fusion module, wherein the loss of the generator, the loss of the discriminator and the classification label loss are respectively updated to update the parameters of the generator, the parameters of the discriminator and the parameters of other modules during the training process.

Description

Technical Field

[0001] This invention relates to the field of sleep quality and sleep disorder diagnosis, specifically to the technical field of sleep classification based on electroencephalogram (EEG) signals, and more specifically to a single-channel EEG sleep classification model and construction method based on adversarial generative network enhancement for a universal environment. Background Technology

[0002] Sleep monitoring and staging play a crucial role in human health and disease diagnosis, providing a wealth of information about individual sleep quality and brain and nervous system-related diseases. Currently, two main sleep stage classification standards are widely used: the Rechtschaffen & Kales (R&K) standard and the American Academy of Sleep Medicine (AAS M) standard. The R&K standard, first proposed in 1968, is based on multimodal physiological temporal signals from polysomnography (PSG), such as electroencephalography (EEG), electrocardiography (ECG), electrooculography (EOG), and electromyography (EMG). PSG divides sleep into seven stages: wake, rapid eye movement (REM), and four non-rapid eye movement (NREM) stages. The NREM stages include stages 1 (S1), 2 (S2), 3 (S3), and 4 (S4). The first two stages are classified as light sleep, while the latter two are called deep sleep or slow-wave sleep. In 2007, the AASM (Audio-Associated Sleep and Respiratory Syndrome) improved the R&K (Real-Wave and Sleep) standard, merging stages S3 and S4 into stage S3, and renaming stages S1, S2, and S3 as N1, N2, and N3. The improved AASM standard divides the sleep cycle into five sleep stages: Wake, REM, N1, N2, and N3. These internationally recognized sleep stage classification standards provide a foundation and consistency for sleep monitoring. Sleep stage classifications based on these standards play an indispensable role in many fields, including healthcare and disease diagnosis, providing crucial support for human health and disease management.

[0003] Traditional EEG sleep stage classification methods typically analyze raw signals using time-frequency and statistical techniques. While these methods have achieved some success, they rely heavily on expert prior knowledge and manual analysis, resulting in a large workload and limited classification accuracy. With technological advancements, machine learning and deep learning-based sleep stage classification methods have simplified model training and achieved significant results, particularly those utilizing multi-channel and multimodal PSG signals. However, wearing multi-channel PSG or EEG sleep monitoring devices is uncomfortable for subjects and can affect sleep quality, thus limiting their application in common environments such as the home. Furthermore, the large number of electrodes makes the devices bulky and inconvenient to carry, and signal bias can occur due to discomfort during wear and human interference during sleep.

[0004] To overcome these challenges, several suitable devices and methods have been developed. For example, sleep monitoring methods based on single-channel EEG use specially designed single-channel EEG devices to sense sleep data and leverage machine learning algorithms to classify sleep stages. Commonly used channels include the left central C3-A2 channel, the right central C4-A1 channel, and channels near the ear in the International 10-20 system, as these acquisition locations often coincide with the subject's comfortable sleeping posture. Based on this, in-ear EEG devices have been developed and used commercially for driver drowsiness monitoring. Furthermore, researchers are exploring sleep monitoring devices based on smartwatches, smartphones, and smart eye masks. However, single-channel data, due to its limited ability to sense only localized information, contains limited information. Simply training models based solely on single-channel data for sleep stage classification can lead to insufficient accuracy. To address this issue, several methods exist for expanding EEG data. For example, patent application CN202310259223.9 provides a neurofeedback modulation method and system for rehabilitation training of patients with brain disorders. This method translates EEG sequence signals into ideal EMG sequence signals and compares these ideal signals with real EMG signals to determine the degree of brain control over EMG in patients with brain disorders. Another example is patent application CN202210071629.X, which provides a cross-subject rehabilitation training method for stroke patients. This method uses motor imagery EEG data from healthy individuals to generate new EEG data, thus expanding the data and using a large amount of expanded healthy EEG data to assist in the diagnosis and rehabilitation of stroke patients.

[0005] Therefore, fully utilizing multi-channel information is crucial for sleep monitoring. However, existing wearable devices, such as wristbands, can only sense limb movement information but cannot capture the true state of the brain. Single-channel devices (such as single-channel in-ear EEG) have limited signal acquisition channels, allowing them to sense only limited information from specific brain locations. Some researchers have attempted to generate ideal electromyography (EMG) data from EEG data and assess the recovery of patients with brain disorders by comparing ideal EMG data with real EMG data. However, this method faces several problems, including the non-stationarity of physiological signals and subject variability, making it unreliable to directly generate corresponding time-series data from raw time-series data. Furthermore, while the generated data can be used to expand and enhance the original data, it is unreliable to use it directly as a basis for judgment. The sleep monitoring effects based on the above-mentioned schemes are far inferior to those based on multi-channel clinical devices. Therefore, fully utilizing the rich information of multi-channel data to make the sleep stage classification effect based on single-channel data approach that based on multi-channel data is a significant challenge. Establishing a mapping between single-channel and multi-channel information is key to solving this challenge. Summary of the Invention

[0006] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a single-channel EEG sleep classification model and its construction method.

[0007] According to a first aspect of the present invention, a method for constructing a single-channel EEG sleep classification model is provided. The method includes: S1, acquiring raw PSG data, wherein the PSG data contains multiple records, each record contains multiple channel data, and each record corresponds to a sleep stage classification label; S2, using a preset EEG channel as the target single channel and multiple other channels as associated channels, selecting target single-channel data and raw associated channel data from each record in the raw PSG data to form a training dataset; S3, constructing an initial classification model, wherein the initial model includes a feature extraction module, a generative adversarial unit, a classification unit, and a fusion module, wherein: the feature extraction module is used to extract features of the target single-channel data and features of multiple raw associated channel data; the generative adversarial unit is used to generate features of multiple associated channels based on the features of the target single-channel data, and includes multiple generative adversarial modules, one generative adversarial module corresponding to one associated channel, and each generative adversarial module... The system includes a generator and a corresponding discriminator. The generator generates corresponding associated channel features based on the features of the target single-channel data. The discriminator identifies the authenticity of the associated channel features generated by the corresponding generator based on the features of the original associated channel data. The classification unit performs sleep stage classification prediction based on the features of the target single-channel data and the features of the multiple associated channels generated therefrom to obtain multiple sleep stage classification prediction results. The fusion module fuses the sleep stage classification prediction results obtained by the classification unit to obtain the final sleep classification result corresponding to the target single-channel data. S4: The initial classification model is trained using the training dataset until convergence to obtain a model consisting of a feature extraction module, multiple generators, a classification unit, and a fusion module. During training, the parameters of the corresponding generator are updated using the loss of each generator, the parameters of the corresponding discriminator are updated using the loss of each discriminator, and the parameters of the feature extraction module, the classification module, and the fusion module are updated using the classification label loss.

[0008] Preferably, the target single-channel feature extraction module is configured as a FeatureNet model.

[0009] Preferably, the classification unit includes multiple classification modules, one of which is used to classify sleep stages based on the associated channel features generated by a generator.

[0010] Preferably, the preset EEG channel is the C3-A2 channel or the C4-A1 channel.

[0011] Preferably, the other multiple channels are multiple channels other than the target single channel, such as EEG channels, ECG channels, EMG channels, and EEG channels.

[0012] Preferably, the other multiple channels are two EEG channels, one ECG channel, two EOG channels, and one EMG channel, which are other EEG channels besides the target single channel.

[0013] Preferably, the loss of each generator is:

[0014]

[0015] Where L(G) represents the generator loss, E represents the expectation, and p z Let z represent the corresponding associated channel feature generated, G(z) represent the corresponding associated channel feature generated by the generator, and D(G(z)) represent the discriminator's discrimination result on the associated channel feature generated by the generator.

[0016] Preferably, the loss of each discriminator is:

[0017]

[0018] Where L(D, G) represents the discriminator loss, p x Let x represent the original associated channel data feature, and D(x) represent the discrimination result of the discriminator on the original associated channel data feature.

[0019] Preferably, the classification label loss is:

[0020]

[0021] Where I represents the number of data samples in a single target channel, N represents the number of generative adversarial modules, and L mc (*) represents the cross-entropy loss, f i t This represents the associated channel feature generated by the generator in the t-th generative adversarial module for the i-th data sample in a single target channel. Indicates the relationship with f i t Predicted classification labels for classification prediction. This represents the actual classification label corresponding to the i-th data sample in the target single channel.

[0022] According to a second aspect of the present invention, a single-channel EEG data sleep classification model based on the method described in the first aspect of the present invention is provided. The model includes: a feature extraction module for extracting features of target single-channel data; multiple generators, each generator corresponding to an associated channel, and each generator for generating corresponding associated channel features based on the features of the target single-channel data; a classification unit for performing sleep stage classification prediction based on the features of the target single-channel data and the features of the multiple associated channels generated therefrom to obtain multiple sleep stage classification prediction results; and a fusion module for fusing the sleep stage classification prediction results obtained by the classification unit to obtain the final sleep classification result corresponding to the target single-channel data.

[0023] Compared with existing technologies, this invention utilizes a GAN generator to generate neural network features for other channels based on existing single-channel EEG neural network features, and uses multiple discriminators to perform one-to-one discrimination for each channel, helping the generator to achieve accurate cross-channel mapping generation. This achieves the goal of obtaining multi-channel information based on single-channel features. The generated multi-channel feature data and the original real feature data are synchronously fused and used, which can enhance the sleep stage classification effect in a universal environment based on single-channel EEG. Attached Figure Description

[0024] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0025] Figure 1 This is a schematic diagram of a method for constructing a single-channel EEG sleep classification model according to an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of the initial classification model structure according to an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is further described in detail below through specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] As described in the background section, the most significant challenge facing existing sleep stage classification schemes based on single-channel data is how to construct a mapping between single-channel information and multi-channel data, thereby acquiring multi-channel information through single-channel data collection to build a better multi-channel classification model based on single-channel data. The purpose of this invention is to fully utilize existing multi-channel data to train models, improving their performance in inference using single-channel data in general environments. Cross-channel mapping is crucial in this process. The inventors have drawn inspiration from the principles of Generative Adversarial Networks (GANs) and applied them to cross-channel mapping, enabling the acquisition of a good multi-channel classification model based on single-channel EEG data, and allowing the model to be applied to general domains. The generator and discriminator components of GANs provide valuable insights into cross-channel mapping and generation.

[0029] In their research on improving sleep stage classification based on single-channel EEG signals using multi-channel data, the inventors encountered difficulties in constructing models for training on multi-channel data and testing on single-channel data, leading to inconsistencies between training and testing. Furthermore, due to the highly complex relationships between channels, constructing a cross-channel correlation network model capable of transferring multi-channel information to single-channel data is extremely challenging. Generative Adversarial Networks (GANs) provide a framework for correlation mapping. By using the specialized generator and discriminator structures of GANs, the model can autonomously learn and determine the mapping relationships between the hidden layer representations of physiological signals in the neural network. For example, the discriminator can assess the reliability of the generated data and its corresponding channel. To avoid mutual interference, this invention considers setting up a dedicated discriminator for each channel. In addition, the inventors found that directly generating physiological time-series signal data is very difficult. Therefore, they decided to utilize the feature extraction capabilities of neural networks to generate data from pre-extracted hidden layer features. This method effectively solves the problem of directly generating physiological time-series signal data.

[0030] In summary, according to one embodiment of the present invention, such as Figure 1 As shown, this invention provides a method for constructing a single-channel EEG sleep classification model. The method includes: S1, acquiring raw PSG data, wherein the PSG data contains multiple records, each record contains multiple channel data, and each record corresponds to a sleep stage classification label; S2, using a preset EEG channel as the target single channel and multiple other channels as associated channels, selecting the target single channel data and the original associated channel data from each record in the raw PSG data to form a training dataset; S3, constructing an initial classification model, according to an embodiment of this invention, as follows... Figure 2As shown, the initial model includes a feature extraction module, a generative adversarial unit (GAN), a classification unit, and a fusion module. The feature extraction module extracts features from the target single-channel data and features from multiple original associated channel data. The GAN generates features for multiple associated channels based on the features of the target single-channel data. It includes multiple GAN modules, each corresponding to one associated channel. Each GAN module includes a generator and a corresponding discriminator. The generator generates corresponding associated channel features based on the features of the target single-channel data, and the discriminator identifies the authenticity of the associated channel features generated by its corresponding generator based on the features of the corresponding original associated channel data. The classification unit is used to perform sleep stage classification prediction based on the features of the target single-channel data and the features of multiple associated channels generated based on it to obtain multiple sleep stage classification prediction results; the fusion module is used to fuse the sleep stage classification prediction results obtained by the classification unit to obtain the final sleep classification result corresponding to the target single-channel data; S4, the initial classification model is trained with the training dataset until convergence to obtain a model consisting of a feature extraction module, multiple generators, a classification unit, and a fusion module. During the training process, the parameters of the corresponding generator are updated using the loss of each generator, the parameters of the corresponding discriminator are updated using the loss of each discriminator, and the parameters of the feature extraction module, the classification module, and the fusion module are updated using the classification label loss.

[0031] The discriminator plays a dual role in this invention: firstly, through adversarial training, it makes the other channel features generated by the generator based on single-channel features more realistic and reliable; secondly, the discriminator can perform fine-grained discrimination of the channels of the generated features, thereby determining their affiliation. This invention also employs a deep learning network-based method, generating new channel features by pre-extracting hidden layer features from the neural network of single-channel data, and establishing mapping relationships between channels using a GAN discriminator. This not only increases the number of data channels but also maintains the original, realistic signal as the primary basis. It not only expands the amount of EEG signal data but also expands multiple channels simultaneously. The new channels maintain the same timeline as the original channels and can also include other modalities besides EEG (such as EMG and ECG), thereby achieving the goal of enhancing the single-channel sleep stage classification effect by generating features from different PSG channels.

[0032] To better understand the present invention, the following embodiments will be presented in turn from multiple aspects to illustrate the solution of the present invention.

[0033] 1. Problem Definition

[0034] Sleep stage classification: Polysomnography (PSG) signals include multi-channel EEG, ECG, EMG, and EOG signals, providing valuable information about sleep stages. To improve the accuracy and consistency of sleep stage classification, these signals are typically divided into multiple 30-second periods. The AASM standard classifies each sleep stage into Wake, REM, N1, N2, and N3 stages. Sleep stage classification models are generally configured to be cross-subject task-based, meaning the training and test sets come from different subjects.

[0035] Generative Adversarial Networks (GANs): G is the generator, D is the discriminator, E represents the expectation, x represents the features of the real data, and z represents the features of the generated data.

[0036] 2. Cross-channel mapping generation method

[0037] This invention utilizes a neural network model for feature pre-extraction and proposes a generator network to generate embedded features for missing channels. Specifically, the system, based on features pre-extracted from channels C3-A2, synchronously simulates and generates features from other channels through the generator network. The main task of multiple discriminators is to distinguish between real and generated data. The discriminators consider features extracted from the original data of other channels as real, while considering features generated from channels C3-A2 as fake, helping the generator generate more believable data. Furthermore, since each new channel has its own corresponding discriminator, the discriminator can simultaneously learn and determine channel attribution information autonomously, helping the generator generate data that more closely resembles a particular channel. In this way, the system can more accurately simulate the features of each other channel, thus achieving a one-to-one mapping between multiple channels. Through an adversarial iterative training process, the generator's goal is to deceive the discriminators, making them unable to distinguish between generated features and features from real channel data. The training process converges when the generated features successfully deceive the discriminators, making them believe that the generated features originate from the real original data of a specific channel. In summary, the method of this invention utilizes a neural network model for feature pre-extraction and achieves feature generation from the C3-A2 channels to other channels through adversarial training of the generator network and discriminator. This method provides a reliable way to process data with missing channels and offers more comprehensive information for EEG signal analysis.

[0038] 2.1 Effective Characterization of Single-Channel Physiological Signals

[0039] To better extract pre-features from single-channel signals, this invention employs a classic dual-channel model called FeatureNet as its backbone. The FeatureNet model consists of two parallel convolutional neural network (CNN) modules with different kernel sizes for extracting multi-resolution features. Each CNN module comprises five convolutional sequences. Each convolutional sequence consists of a one-dimensional CNN layer, a batchnorm layer, a ReLU regularization layer, and a max-pooling layer. The outputs of these two CNN modules are flattened and concatenated as pre-extracted features. Since the basic principles of the FeatureNet model are known to those skilled in the art, the model itself will not be described in detail in this embodiment.

[0040] 2.2 Generation of Pairwise Mappings Across Channels

[0041] This invention uses a neural network model for feature pre-extraction and introduces a generator network to generate embedded features for missing channels. The invention designs multiple generative adversarial modules, each including a generator and a discriminator. Through one-to-one discrimination, the generator generates features for other channels from features of a single EEG channel, establishing a one-to-one mapping generation network. Furthermore, the adversarial training of the discriminator in conjunction with the generator ensures the reliability of the features generated by the generator. Based on the pre-extracted C3-A2 channel features, the generator network synchronously generates features within other channels. The discriminator distinguishes between real and generated data, comparing the real features of other channels with the features generated from C3-A2 channels. Multiple discriminators sequentially assess the category and authenticity of the generated features, achieving a one-to-one reliable association mapping. Through adversarial training, the generator aims to deceive the discriminator, making it unable to distinguish between the generated features and the features of the real channel data. When the training process converges, the generated features successfully deceive the discriminator, making it believe that the generated features originate from the real original data of a specific channel. In summary, this invention utilizes a neural network model for feature pre-extraction and achieves feature generation from the C3-A2 channels to other channels through adversarial training of the generator network and discriminator. This method reliably processes data with missing channels and provides more comprehensive information for EEG signal analysis.

[0042] Discriminator loss formula:

[0043]

[0044] Where p x Let x represent the original associated channel data features, i.e., the true single-channel EEG features, and p zLet z be the corresponding associated channel feature generated, i.e., other channel features generated. D(G(z)) represents the discrimination result of the discriminator on the associated channel features generated by the generator. The discriminator hopes to distinguish between real data and generated data and their corresponding channel affiliation. Therefore, the expected value E of real data feature x is large, while the expected value E of generated data feature z is small.

[0045] Generator loss formula:

[0046]

[0047] The generator aims to be recognized as real data by the discriminator as much as possible, i.e., D(G(z)) is large enough and the expectation of Equation 2 is small enough.

[0048] The discriminator and generator for each channel undergo iterative adversarial training. The training process converges when the features generated by the generator can fool the discriminator into believing that the generated features come from the original data of that specific channel. After convergence, a generator model that can generate features from other channels from a single-channel EEG mapping is obtained. Based on this, the data for each channel is augmented simultaneously to enhance the classification performance of sleep stages.

[0049] 2.3 Multichannel sleep stage classification

[0050] To protect multi-channel data information and ensure the classification accuracy of the model, as mentioned above, this invention designs a naive model that directly extracts and classifies features on a single channel without participating in feature distribution alignment. Furthermore, N generators are constructed for feature generation. During inference, sleep stage classification is achieved by fusing the output decisions from all N generators using data from the single channel. Each single-channel feature and the generated features from other channels will have a predicted classification result, which is a sleep stage classification probability. During fusion, the one with the highest probability is used as the final sleep stage classification. The classification unit is generally implemented using a fully connected deep learning model, that is, using the original real single-channel features and the generated features from other channels, and training them together through a fully connected deep learning model to predict sleep stage classification. Since this classification method is known to those skilled in the art, this invention will not elaborate on the classification unit.

[0051] 3. Experimental Evaluation

[0052] To better verify the effectiveness of the present invention, the present invention was tested on multiple datasets.

[0053] 3.1 Dataset and Preprocessing

[0054] As shown in Table 1, three datasets were used in the experiments of this invention: ISRUC-S1 and ISRUC-S3 (two common subsets of the ISRUC-Sleep database) and the Sleep Heart Health Study (SHHS1) dataset. Among them:

[0055] The ISRUC-S3 dataset contains 10 healthy adults (nine men and one woman, aged between 30 and 58 years).

[0056] The ISRUC-S1 dataset contains 100 adults with sleep disorders (55 men and 45 women, aged 20 to 85).

[0057] SHHS1 dataset: This invention selected 329 subjects with regular sleep patterns based on the apnea-hypopnea index (AHI) and set the signal sampling frequency to 125Hz.

[0058] Table 1

[0059]

[0060] To prepare the data, PSG records were divided into 30-second intervals and then annotated by two experts according to the AASM standard. The first two datasets, ISRUC-S1 and ISRUC-S3, share common characteristics in their composition; each record includes six EEG channels (C3-A2, C4-A1, F3-A2, F4-A1, O1-A2, and O2-A1), two EOG channels (LOC-A2 and ROC-A1), three electromyography channels (chin EMG, left leg movement, and right leg movement), and one electrocardiogram channel. This invention selected all six EEG channels from the ISRUC-Sleep data to explore cross-EEG channel associations and resampled the signals at a frequency of 100Hz. From the 14 channels of the SHHS1 dataset, this invention selected the six most commonly used time-series physiological signal channels in sleep stage classification tasks (i.e., two EEG channels, one ECG channel, one EMG channel, and two EOG channels) to explore cross-channel mappings of PSG signals. It should be noted that channels C3-A2 and C4-A1 are channels that do not affect the sleep quality of the subjects and can both be used as target single channels, while other channels can be used as associated channels. However, since right lateral decubitus is a more comfortable sleeping position for patients with sleep disorders, the experiment of this invention uses channel C3-A2, which is not compressed at all in this sleeping position and is often used for sleep stage classification, as the target single channel EEG.

[0061] 3.2 Evaluation Indicators

[0062] Accuracy, F1 score (F1), Macro F1, and Kappa are the evaluation metrics for this work. Their definitions are as follows:

[0063]

[0064]

[0065]

[0066]

[0067] Where TP represents the number of samples correctly classified into the current sleep stage, FP represents the number of samples from other sleep stages that were misclassified into the current sleep stage, FN represents the number of samples from the current sleep stage that were misclassified into other stages, and TN represents the number of samples from other sleep stages that were accurately classified as not belonging to the current sleep stage. The F1 score is defined as the harmonic mean of Recall (measures how many samples that actually belong to a specific sleep stage are successfully predicted by the model) and Precision (measures how many samples that are predicted to belong to a specific sleep stage actually belong to that specific sleep stage). r represents the specific sleep stage category, and R represents the total number of sleep stage categories. p o This refers to the relative observational consistency among evaluators, while p e It represents the assumed probability of consistency of chance or consistency of probability.

[0068] 3.3 Comparison Methods

[0069] FeatureNet is a classic sleep stage classification model composed of dual-scale CNNs. For consistency, it was also used as the feature pre-extraction network in the experiments on the three datasets in this paper. In the comparative method of this invention, the performance of the method that classifies using only the features extracted by the FeatureNet model (represented by FeatureNet), the method that classifies using features extracted by the FeatureNet model, and the method that classifies using features generated from those features (represented by GAN) on each dataset is compared. In the comparison process, this invention uses 10-fold cross-validation to split the training and test sets, that is, the ratio of the training set to the test set in this invention is 9:1. Then, 20% of the training set is randomly selected as the validation set, the best model validated on the validation set is saved, and the model is tested on a completely unseen new user test set. The results are shown in Table 2.

[0070] Table 2

[0071]

[0072] The experimental results show that using the method of this invention to perform multi-channel synchronous data augmentation on the dataset results in a significant improvement compared to the FeatureNet method without multi-channel synchronous data augmentation.

[0073] Compared with existing technologies, this invention utilizes a GAN generator to generate neural network features for other channels based on existing single-channel EEG neural network features, and uses multiple discriminators to perform one-to-one discrimination for each channel, helping the generator to achieve accurate cross-channel mapping generation. This achieves the goal of obtaining multi-channel information based on single-channel features. The generated multi-channel feature data and the original real feature data are synchronously fused and used, which can enhance the sleep stage classification effect in a universal environment based on single-channel EEG.

[0074] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order, as long as the required function can be achieved.

[0075] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0076] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can be, for example, including but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0077] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A single-channel electroencephalogram sleep classification model construction method, characterized in that, The method comprises: S1, acquiring original PSG data, the PSG data containing a plurality of records, each record containing a plurality of channel data, and each record corresponding to a sleep stage classification label; S2, taking a preset electroencephalogram channel as a target single channel, and taking a plurality of other channels as associated channels, to filter out target single channel data and original associated channel data from each record in the original PSG data to form a training data set; wherein the plurality of other channels are a plurality of channels selected from the group consisting of other electroencephalogram channels, electrocardiogram channels, electromyogram channels and electrooculogram channels other than the target single channel; S3, constructing an initial classification model, the initial model comprising a feature extraction module, a generative adversarial unit, a classification unit and a fusion module, wherein: the feature extraction module is configured to extract features of the target single channel data and features of the plurality of original associated channel data; the generative adversarial unit is configured to generate features of the plurality of associated channels based on the features of the target single channel data, and comprises a plurality of generative adversarial modules, one generative adversarial module corresponding to one associated channel, and each generative adversarial module comprising a generator and a corresponding discriminator, the generator being configured to generate corresponding associated channel features based on the features of the target single channel data, and the discriminator being configured to distinguish whether the associated channel features generated by the corresponding generator are true or false based on the features of the corresponding original associated channel data; the classification unit is configured to respectively perform sleep stage classification prediction based on the features of the target single channel data and the features of the plurality of associated channels generated based thereon to obtain a plurality of sleep stage classification prediction results; the fusion module is configured to fuse the sleep stage classification prediction results obtained by the classification unit to obtain a final sleep classification result corresponding to the target single channel data; S4, training the initial classification model to convergence using the training data set to obtain a model comprising the feature extraction module, the plurality of generators, the classification unit and the fusion module, wherein the parameters of each generator are updated using the loss of each generator, the parameters of each discriminator are updated using the loss of each discriminator, and the parameters of the feature extraction module, the classification module and the fusion module are updated using the classification label loss.

2. The method of claim 1, wherein, The target single channel feature extraction module is configured as a FeatureNet model.

3. The method of claim 1, wherein, The classification unit comprises a plurality of classification modules, and one classification module is configured to perform sleep stage classification based on the associated channel features generated by one generator.

4. The method of claim 3, wherein, The preset electroencephalogram channel is a C3-A2 channel or a C4-A1 channel.

5. The method of claim 4, wherein, The plurality of other channels are 2 electroencephalogram channels, 1 electrocardiogram channel, 2 electrooculogram channels and 1 electromyogram channel selected from the group consisting of other electroencephalogram channels other than the target single channel.

6. The method of claim 1, wherein, The loss of each generator is: wherein, denotes the generator loss, denotes the expectation, denotes is a corresponding associated channel feature generated by the generator, denotes the discrimination result of the discriminator on the associated channel feature generated by the generator.

7. The method of claim 6, wherein, The loss of each discriminator is: wherein, represents a discriminator loss, represents is an original associated channel data feature, represents a discrimination result of the discriminator on the original associated channel data feature.

8. The method of claim 1, wherein, The classification label loss is: wherein, represents a quantity of data samples of the target single-channel data, represents a number of generative adversarial modules, represents a cross-entropy loss, represents a target single-channel first data sample in a first generative adversarial module, represents a prediction classification label for classification prediction, represents an actual classification label corresponding to the target single-channel first data sample.

9. A single-channel electroencephalogram data sleep classification model based on the method of any one of claims 1-8, characterized in that, The model comprises: a feature extraction module configured to extract features of the target single channel data; a plurality of generators, each generator corresponding to one associated channel, and each generator being configured to generate corresponding associated channel features based on the features of the target single channel data; The classification unit is configured to respectively perform sleep stage classification prediction according to features of the target single-channel data and features of a plurality of associated channels generated based on the target single-channel data to obtain a plurality of sleep stage classification prediction results. The fusion module is configured to fuse the sleep stage classification prediction results obtained by the classification unit to obtain a final sleep classification result corresponding to the target single-channel data.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer readable medium, and the computer program is executable by a processor to implement the steps of the method of any one of claims 1 to 8.

11. An electronic device, comprising: The computer program comprises: one or more processors; a memory device storing one or more programs that, when executed by the one or more processors, cause the electronic device to implement the steps of the method of any one of claims 1 to 8.

Images