Single-channel electroencephalogram sleep classification model construction method and model

Abstract

The application provides a single-channel electroencephalogram sleep classification model construction method, which comprises the following steps: S1, obtaining original PSG data, wherein the PSG data comprises a plurality of records, each record comprises a plurality of channel data, and each record corresponds to a sleep stage classification label; S2, screening 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 a preset electroencephalogram channel is taken as the target single channel, and a plurality of other channels are taken as the associated channels; S3, constructing an initial classification model, wherein the initial model comprises an initial associated feature mining unit, a classification unit and a fusion module; and S4, training the initial classification model to convergence by using the training data set, wherein in the training process, the model parameters are updated by using a comprehensive loss which is constructed based on the distance loss between the associated channel features and the features of the associated channel data and the classification label loss.

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 method and model for constructing a single-channel EEG sleep classification model for a universal environment based on cross-channel association mining. Background Technology

[0002] Sleep monitoring and staging are crucial for humans, providing valuable data for assessing individual sleep quality and diagnosing sleep-related disorders. In particular, sleep stage classification has wide applications in disease monitoring, planning, and intervention, and is of great significance to human health.

[0003] Currently, there are two main standards for classifying sleep stages: the Rechtschaffen & Kales (R&K) standard and the American Academy of Sleep Medicine (AASM) standard. The R&K standard, first proposed in 1968, divides sleep into seven stages based on multimodal physiological time-series signals from polysomnography (PSG), including electroencephalography (EEG), electrocardiography (ECG), electrooculography (EOG), and electromyography (EMG)—namely, wakefulness, rapid eye movement (REM), four non-rapid eye movement (NREM) stages, and a motor phase. NREM comprises four stages, named Stage 1 (S1), Stage 2 (S2), Stage 3 (S3), and Stage 4 (S4). The first two stages belong to light sleep, while the latter two belong to deep sleep, also known as slow-wave sleep. In 2007, the AASM improved the R&K standard, merging stages S3 and S4 into a single stage S3, and renaming stages S1, S2, and S3 as N1, N2, and N3, respectively. The revised AASM standard divides the sleep cycle into five stages: Wake, REM, N1, N2, and N3, corresponding to the five categories of sleep stage classification. Based on these internationally recognized sleep stage classification standards, sleep monitoring plays an indispensable role in many fields such as healthcare and disease diagnosis.

[0004] There are two main types of sleep stage classification methods under current technology: one is the traditional EEG sleep stage classification method, and the other is the sleep stage classification method based on machine learning and deep learning.

[0005] Traditional EEG sleep stage classification methods typically employ time-frequency analysis of raw signals and statistical methods to analyze the results. While these traditional methods have yielded some results, they heavily rely on expert prior knowledge and manual analysis, making them labor-intensive and limiting their classification accuracy.

[0006] Fortunately, sleep stage classification based on machine learning and deep learning has simplified the training process of classification models and achieved considerable results. In particular, classification models based on multi-channel, multi-modal PSG signals for sleep stage classification have shown excellent results. However, wearing sleep monitoring devices with multi-channel PSG or EEG is uncomfortable, affecting the sleep quality of subjects and limiting the application of such devices in general environments (i.e., home environments). Multiple bulky electrodes lack portability and cause signal deviations due to discomfort during wear and manual compression during sleep. To overcome this challenge, several universal devices and methods have been developed. Sleep monitoring methods based on single-channel EEG use specially designed single-channel EEG devices to sense sleep data and use machine learning algorithms to classify sleep stages. Commonly used channels are the C3-A2 channel located in the left center and the C4-A1 channel in the right center, as well as channels near the ear, in the international 10-20 system, because these channels allow subjects to lie in a comfortable position. Based on this, researchers have developed in-ear EEG devices for commercial driver drowsiness monitoring. In addition, researchers have explored sleep monitoring devices based on smartwatches / bands, smartphones, and smart eye masks. To fully utilize existing multi-channel data, it is essential to train models that can use single-channel data for inference applications in a universal environment and to improve single-channel EEG sleep stage classification by leveraging cross-channel correlations. With existing technology, some researchers have proposed mask-based pre-training fine-tuning methods to facilitate single-channel EEG models actively learning information from other channels, or to utilize coupled learning to reveal the consistency and dissimilarity between EEG signals and other modalities, thereby enhancing the performance of EEG-based models. Furthermore, other researchers have proposed machine learning algorithms for sleep staging using prefrontal single-channel EEG signals, performing decomposition and filtering based on signal frequency, and then combining this with the XGBoost machine learning method for sleep stage classification.

[0007] It's important to note that while non-brain monitoring devices like wristbands can sense limb movement, they cannot capture the true state of the brain. In-ear EEG devices, with their limited signal acquisition channels, mean that sleep monitoring devices based on single-channel EEG can only perceive limited information from specific brain locations. This places high demands on computational algorithms to fully extract effective information. For example, multi-view models can be used to extract information from single-channel EEG from multiple angles, potentially improving performance. However, without clear data and label guidance, the training and learning process may be randomly and blindly biased. Furthermore, existing multi-view methods are not designed for universal sleep monitoring tasks and have limitations. Applying pre-training and fine-tuning paradigms requires collecting large amounts of data and consuming significant computational resources; the fusion of multimodal information cannot be applied to universal monitoring environments. Additionally, while existing single-channel prefrontal EEG data acquisition and wearing are convenient, the cumbersome and complex preprocessing process, along with the relatively low classification accuracy of corresponding machine learning methods, leads to limitations. Therefore, a more effective solution for accurate sleep stage classification based on single-channel data is urgently needed. Summary of the Invention

[0008] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for constructing a single-channel EEG sleep classification model.

[0009] 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 an initial association feature mining unit, a classification unit, and a fusion module, wherein: the initial association mining unit is used to extract features from the target single-channel data and mine features from multiple associated channels based on the features of the target single-channel data, and includes a target single-channel feature extraction module and multiple initial associations. The feature mining module includes an initial association feature mining module, each of which comprises an original association channel data feature extraction module and an association feature mining module. The classification unit is used to perform sleep stage classification prediction based on the target single-channel feature and the features of multiple association channels mined from 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 using the training dataset until convergence to obtain a model composed of the target single-channel feature mining module, the association feature mining module, the classification unit, and the fusion module. During the training process, the model parameters are updated using a comprehensive loss constructed from the distance loss between the association channel features mined from the target single-channel data and the features of the association channel data, and the classification label loss.

[0010] Preferably, the target single-channel feature extraction module is configured as a FeatureNet model.

[0011] Preferably, each original associated channel data feature extraction module is used to extract features of its corresponding original associated channel data, and each associated feature mining module is used to mine features of its corresponding associated channel based on the features of the target single channel data.

[0012] Preferably, the classification unit includes multiple classification modules, one of which is used to classify sleep stages based on the target single-channel features or the features of the mined associated channels.

[0013] Preferably, the preset EEG channel is the C3-A2 channel or the C4-A1 channel.

[0014] 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.

[0015] 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.

[0016] Preferably, the overall loss is:

[0017]

[0018] in, Indicates the loss of classification labels. It is a distance loss, and:

[0019]

[0020] Where I represents the data sample size of the target single-channel data, N represents the number of initial correlation mining modules, 2N represents the total number of original correlation channel data feature extraction modules and correlation feature mining modules corresponding to all initial correlation mining modules in the correlation mining unit, and L mc (*) represents the cross-entropy loss, f i t This represents the feature obtained by the i-th data sample in the target single channel in the t-th module. This represents the predicted classification label based on the features obtained from the i-th data sample in the t-th module of the target single channel. This represents the actual classification label corresponding to the i-th data sample in the target single channel;

[0021]

[0022] in, Let α represent the maximum average difference loss between the features of the original associated channel data mined by the initial management feature mining module in the c-th initial association mining module and the features of the associated channels mined by the associated post mining module based on the target single-channel features. α is the weight of the loss of the c-th initial association mining module.

[0023]

[0024] Where k(*) represents the kernel function that maps the data to a high-dimensional space, f i n f represents the feature of the i-th data sample in the target single channel. i n′ This represents the features of the associated channels mined by the association mining module in the c-th initial association mining module based on the features of the i-th data sample in the target single channel. This indicates that the association mining module in the c-th initial association mining module mines the association channel features based on the i′-th original association channel data corresponding to the i-th data sample of the target single channel.

[0025] According to a second aspect of the present invention, a sleep classification model for single-channel EEG data is provided. The model includes: an association feature mining unit, comprising a target single-channel feature extraction module and multiple association feature mining modules, wherein the target single-channel feature extraction module is used to extract features from the target single-channel EEG data, and each association feature mining module is used to mine features of an association channel based on the features of the target single-channel EEG data; a classification unit, comprising multiple classification modules, one classification module corresponding to either the target single-channel feature extraction module or the association feature mining module, wherein each classification module is used to perform sleep stage classification prediction based on the features obtained by its corresponding module to obtain multiple sleep stage classification prediction results, wherein each sleep stage classification prediction result is the probability that the current single-channel EEG data belongs to that sleep stage; and a fusion module, used to fuse the sleep classification prediction results of each classification module to obtain the final sleep classification result of the target single-channel data, wherein the sleep classification prediction result with the highest probability is taken as the final sleep stage classification result.

[0026] Compared with existing technologies, the advantages of this invention are as follows: The proposed solution is a sleep stage classification method for universal sleep monitoring scenarios. The system mainly includes three modules: effective representation of single-channel physiological signals, pairwise correlation mining between channels, and multi-view sleep stage classification. The effective representation of single-channel physiological signals is used to extract and fuse multi-scale information of sleep physiological signals, and serves as the backbone model for naive single-channel feature extraction and classification models, as well as cross-channel correlation networks. The pairwise correlation mining model between channels is driven by the fusion of data and intelligent algorithms. Utilizing multi-channel data and paired cross-channel correlation mining algorithms, it transfers multi-channel PSG signal information to single-channel EEG signals, significantly enhancing the sleep stage classification effect in a universal environment based on single-channel EEG. Through multi-view fusion, only a single EEG channel is needed in the inference stage to fuse all information from multi-channel PSG from multiple angles, thereby achieving more effective and universal single-channel sleep stage classification. Attached Figure Description

[0027] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0028] 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;

[0029] Figure 2 This is a schematic diagram of the initial classification model structure according to an embodiment of the present invention;

[0030] Figure 3This diagram illustrates a comparison between the performance of existing classification models and the classification model obtained by this invention.

[0031] Figure 4 This is a schematic diagram illustrating the principle of a single-channel EEG sleep classification model according to an embodiment of the present invention.

[0032] Figure 5 This is a schematic diagram of feature embedding according to an embodiment of the present invention. Detailed Implementation

[0033] 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.

[0034] As described in the background section, the inventors, while researching a multi-view sleep stage classification model based on single-channel EEG signals, discovered that in existing multi-view methods, each view's model, lacking definite data guidance, would autonomously learn in arbitrary directions, resulting in improved classification performance. Through research, the inventors found that by using cross-channel association to feed multi-channel PSG signal information into a single-channel EEG signal model, and through multi-channel data-driven guidance, the intelligent model can be directionally guided to mine across channels. Based on this, the multi-view model's learning process becomes more orderly and effective, and the trained model can perform more accurate sleep stage classification based on single-channel EEG signals. The invention is described in detail below with reference to the accompanying drawings and embodiments.

[0035] 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, comprising: 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 an initial association feature mining unit, a classification unit, and a fusion module. The initial association mining unit is used to extract features from the target single-channel data and mine features from multiple association channels based on the features of the target single-channel data. It includes a target single-channel feature extraction module and multiple initial association feature mining modules. Each initial association feature mining module includes an original association channel data feature extraction module and an association feature mining module. The classification unit is used to perform sleep stage classification prediction based on the target single-channel features and the features of multiple association channels mined from them 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 using the training dataset until convergence to obtain a model composed of the target single-channel feature mining module, the association feature mining module, the classification unit, and the fusion module. During training, the model parameters are updated using a comprehensive loss constructed from the distance loss between the association channel features mined from the target single-channel data and the features of the association channel data, and the classification label loss.

[0036] In summary, the present invention is a cross-channel association mining scheme, referred to as CCAM (hereafter referred to as CCAM for convenience), aimed at improving the accuracy of sleep stage classification based on single-channel EEG. In addition to a naive feature extraction model (i.e., the target single-channel feature extraction module), CCAM utilizes synchronous single-channel EEG signals and other high-density PSG signals to establish pairwise association mining models. These cross-channel association mining models facilitate the transfer of prior information from other relevant channels to the target EEG channel. During the inference phase, CCAM uses the association mining models to extract effective features from the target EEG channel and improves the accuracy of sleep stage classification through multi-view information fusion. The CCAM method uses high-density PSG signals to drive the multi-view model, mining multi-channel information from single-channel EEG signals. To address the challenge of different features and distributions among PSG signal channels, the present invention employs maximum mean difference (MMD) to align the feature distributions between a single EEG signal channel and other PSG signal channels. The final sleep stage classification result is obtained through decision fusion of the multi-view model.

[0037] The following embodiments will be used to describe the solution of the present invention from multiple aspects.

[0038] 1. Problem Definition

[0039] Sleep stage classification: Polysomnography (PSG) signals, including multi-channel EEG, ECG, EMG, and EOG signals, provide 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 (non-REM), N2 (non-REM), and N3 (non-REM). Sleep stage classification models are generally set to be cross-subject task-oriented, meaning the training and test sets come from different subjects. The sleep stage classification problem is defined as y i =G y (G f (x i ), where x represents a sample, i represents the index of the training sample, and G f G represents the feature extractor. y Represents the label classifier, y i This represents the predicted sleep stage classification result for the i-th sample. Specifically, the single-channel sleep stage classification problem can be defined as... Where n∈{0,…,N} represents the channel of the sample, i / j∈{1,…,I / J}, and I and J represent the sample size of the training set and the test set, respectively. This represents the characteristics of channel n in sample i. The softmax value representing the prediction result of channel n comes from sample i. This value is generally a probability value. According to the problem definition, in this invention, sleep stages are classified by single-channel EEG signals and multi-channel features mined based on single-channel EEG signals, and the one with the highest probability is taken as the final sleep stage classification result.

[0040] Cross-channel correlation mining: For a single channel of a PSG, the purpose of cross-channel correlation mining is to represent and model the signal from different perspectives. This invention uses M(f n ) represents the naive feature extraction model, where n represents the channel index. Additionally, M c (f n ,f n′ The first part is a feature extraction model based on association mining, where c∈{1,…,C} is the index of the association model, C=N, and N represents the total number of association channels. n′∈(1,N) and n′≠n represent the indices of other channels. A naive feature extraction model and N association feature extraction models form a multi-view framework with N+1 different perspectives. Feature f n Training is performed using data from a single channel, while the feature f n′ Training is performed using data from N other channels. The sleep stage classification result of multi-view learning can be defined as y = G. y (f0 ,…,f C ).

[0041] 2. Cross-channel correlation mining

[0042] 2.1 Single-channel signal feature extraction

[0043] To better extract information from single-channel signals, this invention employs a classic dual-path convolutional model called FeatureNet as its backbone. The FeatureNet model consists of two parallel convolutional neural network (CNN) modules with different kernel sizes and a fully connected layer. These two CNN modules are used to extract multi-resolution features, and each CNN module consists of five convolutional layers. Each convolutional sequence consists of a one-dimensional CNN layer, a Batchnorm and ReLU regularization layer, and a max-pooling layer. The outputs generated by these two CNN modules are flattened and concatenated, then fed into a fully connected layer for classification. 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.

[0044] In the FeatureNet model, the outputs of two CNN modules are flattened and concatenated together as a connection feature f. n Then it is used for classification.

[0045] Single-channel EEG signals contain limited sleep-related information. Therefore, it is essential to transfer information from multi-channel PSG signals to single-channel EEG-based models using a multi-view cross-channel correlation method. Specifically, such as... Figure 3 As shown, model (a) is a model without cross-channel association mining, and model (b) is a model with cross-channel association mining. In (a), the model blindly tilts in different directions. In (b), the model tilts in a defined direction under the supervision of data from other channels. This tilting process can also be called intelligent modeling and data-driven algorithmic flow. It utilizes high-density PSG signals to drive the model to mine multi-view cross-channel information from single EEG channel data.

[0046] 2.2 Pairwise correlation mining between channels

[0047] The goal of the CCAM method in this invention is to merge information from multiple EEG channels and establish paired cross-channel association models (each pair of cross-channel association models includes a feature extraction module for the original associated channel data and an association feature mining module) to improve the performance of sleep stage classification. In addition to the naive feature extraction model, CCAM utilizes synchronous single-channel EEG signals and other high-density PSG signals to establish paired association mining models. These cross-channel association mining models facilitate the transfer of multi-channel information from other channels to the target EEG channel. During inference, CCAM uses the association mining models to extract effective features from other channels from the target EEG channel.

[0048] Therefore, CCAM is a universal sleep stage classification method that aims to build a cross-channel association mining model to improve the sleep stage classification performance based on single-channel EEG data. The overall framework diagram of the CCAM method is shown below. Figure 4 As shown, effective representation of single-channel physiological signals, cross-channel pairwise correlation mining, and multi-view sleep stage classification are used to extract effective representation of single-channel physiological time-series signals, cross-channel correlation mining, and multi-view fusion, respectively. Among them, the green f... 0 This represents the characteristics obtained from the target single-channel signal; the other f... C This represents the features of other channels obtained through association mining, and then these features (f) 0 ,…,f C The probability predicted by the classification is the final classification result.

[0049] Specifically, during the training process, this invention uses the maximum mean difference (MMD) to measure the distance between cross-channel features mined based on single-channel signal features and actual features of other channels, thereby guiding the training of the mining model.

[0050] The Maximum Mean Dissimilarity (MMD) is a distance metric that encodes distributions using the Reproduction Kernel Hilbert Space (RKHS), quantifying the differences between these encoded distributions into a test statistic for comparison. This invention's CCAM method utilizes the MMD metric to align features from different channels. The goal of this invention is to merge information from multiple EEG channels and establish paired cross-channel association models to improve the performance of sleep stage classification. However, these channels often have unique features and distributions, making the association process challenging. To overcome this challenge, this invention utilizes the MMD metric to align features extracted from a single channel. Features extracted from other channels The distribution of M. c (f n ,f n′By addressing the differences in distributions, this invention ensures that features from a single channel possess richer sleep stage-related information, unique to other channels. This alignment process enables the invention to effectively mine additional channel-related information from the target channel, improving its overall accuracy and reliability. Most importantly, during the inference phase, CCAM effectively enhances the performance using only single-channel data, which is highly advantageous for ubiquitous applications. Pairwise multi-view networks combine features from single channels and other channels. In the pairwise association model, model M... c (f n Features f in ) n By aligning feature distributions, information from other channels is learned, enabling integration and utilization during the inference phase. This multi-view network prevents a single model from being overwhelmed by distributed alignments in various directions.

[0051] Using traditional MMD metrics to measure f n and f n′ The formula is:

[0052]

[0053] Where φ(·) represents the function that maps the feature distribution to RKHS, and k represents the kernel function that maps the data to a high-dimensional space. For the CCAM method of this invention, since f n and f n′ Since the pairs are paired and aligned, with equal sample sizes (p = q), the distance-based loss function of this invention is further simplified as follows:

[0054]

[0055] Where k(*) represents the kernel function that maps the data to a high-dimensional space, f i n f represents the feature of the i-th data sample in the target single channel. i n′ This represents the features of the associated channels mined by the association mining module in the c-th initial association mining module based on the features of the i-th data sample in the target single channel. This indicates that the association mining module in the c-th initial association mining module mines the association channel features based on the i′-th original association channel data corresponding to the i-th data sample of the target single channel.

[0056] The total N pairs of related network losses are summed proportionally:

[0057]

[0058] Where c represents the index of the pairwise feature association network, and α is the weight of the loss for each association network. Loss calculation utilizes the multi-class cross-entropy function. Class loss It is calculated based on the corresponding category labels in all 2N+1 model groups:

[0059]

[0060] in, Let α represent the maximum average difference loss between the features of the original associated channel data mined by the initial management feature mining module in the c-th initial association mining module and the features of the associated channel mined by the associated post mining module based on the target single channel features. α is the weight of the loss of the c-th initial association mining module.

[0061] The total loss is calculated as follows:

[0062]

[0063] By training the cross-channel association mining model to convergence using the total loss, multiple trained association mining modules can be used to mine features of other channels based on single-channel EEG signals.

[0064] 2.3 Multi-view Sleep Stage Classification

[0065] To protect multi-channel data information and ensure the classification accuracy of the model, as mentioned earlier, this invention designs a naive model that directly extracts and classifies features on a single channel without participating in feature distribution alignment. Furthermore, N pairs of pairwise association models are constructed for feature extraction and classification, and the features of a single channel are aligned one-to-one with the features of other channels. During inference, sleep stage classification is achieved by fusing the output decisions of all N+1 models using data from a single channel. These models include N models from the pairwise association models and a single model. (Referring to...) Figure 3 This invention designs a naive model M(f) 0 The system directly extracts and classifies features on a single channel without participating in feature distribution alignment. Furthermore, it constructs N pairwise association models for feature extraction and classification, aligning features from individual channels one-to-one with features from other channels. During inference, sleep stage classification is performed through decision fusion of the outputs of all N+1 models, which utilize cross-channel features mined from single-channel data. These models include N M pairs of pairwise association models. c (f n ) model and naive model M(f 0 ).

[0066] 3. Experimental Evaluation

[0067] To better verify the effectiveness of the present invention, the present invention was tested on multiple datasets.

[0068] 3.1 Data Preprocessing

[0069] 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:

[0070] The ISRUC-S3 dataset contains 10 healthy adults (nine men and one woman, aged between 30 and 58 years).

[0071] The ISRUC-S1 dataset contains 100 adults with sleep disorders (55 men and 45 women, aged 20 to 85).

[0072] 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.

[0073] Table 1

[0074]

[0075] 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-PSG signal channel associations. 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.

[0076] 3.2 Evaluation Indicators

[0077] Accuracy, F1 score (F1), Macro F1, and Kappa are the evaluation metrics for this work. Their definitions are as follows:

[0078]

[0079]

[0080]

[0081]

[0082] 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.

[0083] 3.3 Comparison Methods

[0084] This invention uses FeatureNet as the backbone network and a naive model. Therefore, this invention compares the performance of the model on each channel. The experimental results are shown in Tables 2, 3, and 4. Table 2 shows the experimental results of FeatureNet on single-channel data in the ISRUC-S1 dataset, Table 3 shows the experimental results of FeatureNet on single-channel data in the ISRUC-S3 dataset, and Table 4 shows the experimental results of FeatureNet on single-channel data in the SHHS1 dataset.

[0085] Table 2

[0086]

[0087]

[0088] Table 3

[0089]

[0090] Table 4

[0091]

[0092] The experimental results show that FeatureNet performs best on each channel of the three datasets. On the ISRUC-S1 dataset, the C3-A2 EEG channel performs best, achieving a maximum accuracy of 0.7179, a maximum Macro F1 score of 0.6750, and a maximum Kappa score of 0.6347, all exceeding the average accuracy, Macro F1, and Kappa scores by approximately two, three, and three percentage points, respectively. The second-highest result comes from the F3-A2 EEG channel, which is also located on the left side of the head. On the ISRUC-S3 dataset, the F3-A2 EEG channel performs best, achieving a maximum accuracy of 0.7030, a maximum Macro F1 score of 0.6643, and a maximum Kappa score of 0.6162, all exceeding the average accuracy, Macro F1, and Kappa scores by approximately two percentage points. The second-highest result comes from the 01-A2 EEG channel. On the SHHS1 dataset, the C3-A2 EEG channel performed best, achieving the highest accuracy of 0.7748, the highest Macro F1 score of 0.6330, and the highest Kappa score of 0.6820, all exceeding the average accuracy, Macro F1, and Kappa scores by approximately ten percentage points. The second-highest result came from the C4-A1 EEG channel. Experimental results indicate that the C3-A2 channel performed best in classifying sleep stages on the ISRUC-S1 and SHHS1 datasets, while the F3-A2 channel, also located on the left side of the head, performed best on the ISRUC-S3 dataset.

[0093] To compare the effectiveness of the method of the present invention with other methods, the present invention selects the following existing sleep stage classification methods for comparison:

[0094] FeatureNet: This model is a classic sleep stage classification model composed of dual-scale CNNs. For consistency, it also served as the backbone or naive feature extraction network for the CCAM method and Coupled Learning method in the experiments on the three datasets in this paper.

[0095] AttnSleep: The model structure is based on a multi-resolution convolutional neural network and incorporates a multi-head attention mechanism to capture temporal dependencies.

[0096] DAN: It builds a model structure based on the CNN-Bi-GRU model and uses MMD metric loss to align the distribution of features among users.

[0097] MaskSleepNet: The model architecture is built on a multi-scale convolutional neural network. During data loading, data from channels other than C3-A2 is randomly masked. This encourages the model to actively learn information from these other channels. During inference, only data from the C3-A2 channels is used.

[0098] Coupled Learning: FeatureNet is used to extract initial feature representations for channels C3-A2 and shared initial feature representations for the remaining five channels. During training, common and private representations of these two types are further extracted, and distribution alignment techniques are used to integrate information from other channels into C3-A2. During inference, only the data from channels C3-A2 is used.

[0099] This invention compares CCAM with the methods described above using the same experimental setup. Each comparison method is evaluated using 10-fold cross-validation to split the training and test sets. The ratio of training to test sets in this invention is 9:1. Then, 20% of the training set is randomly selected as the validation set. This invention saves the best-performing model validated on the validation set and tests it on a completely unseen new user test set. Experimental results for different methods on the ISRUC-S1, ISRUC-S3, and SHHS1 datasets are shown in Tables 5, 6, and 7, respectively. DAN also uses MMD as the loss metric, but it addresses the issue of inconsistent data distribution across subjects. The other two are also EEG signal classification methods based on cross-channel association mining. Their focus is on coordinating data distribution across different topics to address topic dependency issues. In contrast, the method in this invention employs a cross-channel association model to address the challenge of aggregating multi-channel information into a single channel, making it universally applicable. MaskSleepNet and coupled learning aim to integrate multi-channel or multimodal information to improve single-channel classification performance. MaskSleepNet utilizes a masking mechanism to enable the model to actively learn from multiple channels, thereby capturing the benefits of multi-channel information. Coupled learning methods further mine cross-channel common and private representations of pre-extracted features. Compared to state-of-the-art methods, the CCAM method of this invention provides more efficient sleep stage classification using single-channel signals on all three datasets.

[0100] Table 5

[0101]

[0102] Table 6

[0103]

[0104]

[0105] Table 7

[0106]

[0107] As shown in Table 5, on the ISRUC-S1 dataset, the CCAM method of this invention achieved the highest Accuracy (0.7365), Macro F1 (0.6895), and Kappa (0.6556), which is about 1 percentage point higher than traditional methods. This state-of-the-art sleep stage classification method outperforms the DAN method using MMD for cross-subject distribution alignment by about 4 percentage points and other cross-channel network construction methods by 7 percentage points. As shown in Table 6, on the ISRUC-S3 dataset, the CCAM method of this invention achieved the highest Accuracy (0.7457), Macro F1 (0.7102), and Kappa (0.6702), which is about 4 percentage points higher than previous results. This state-of-the-art sleep stage classification method outperforms the DAN method using MMD for cross-subject distribution alignment by about 6 percentage points and other cross-channel network construction methods by 6 percentage points. As shown in Table 7, on the SHHS1 dataset, the CCAM method of this invention achieved the highest Accuracy of 0.8056, Macro F1 of 0.6557, and Kappa of 0.7248, which is about 1 percentage point higher than all other methods.

[0108] Experimental results demonstrate that cross-channel information fusion improves sleep stage classification more effectively than the subject dependency problem, and show that the cross-channel association model of this invention is more efficient than other cross-channel association methods. Furthermore, experimental results show that the method of this invention has significant improvements compared to using a single channel, especially when dealing with smaller datasets. This improvement can be attributed to the method of this invention, which effectively utilizes intelligent models and data-driven algorithms. The goal of this invention is to enhance the applicability of sleep stage classification in scenarios with limited data and single channels. As the dataset size increases, the benefits of training with large datasets become more apparent. At this point, there is no need to further enhance the model's performance by integrating multi-channel information. Finding the appropriate balance between data size and accuracy remains an ongoing challenge in the field of deep learning.

[0109] Furthermore, to compare different distance metrics, namely KL divergence, CORAL, CMD, and MMD, to evaluate their performance on the ISRUC-S1 and ISRUC-S3 datasets, the comparison results are summarized in Tables 8 and 9. Among these metrics, MMD achieved the highest accuracy scores on the ISRUC-S1 and ISRUC-S3 datasets, at 0.7365 and 0.7457, respectively. These findings indicate that MMD outperforms other distance metrics in transmitting cross-channel information and building association models. To emphasize the importance of cross-channel information integration in cross-channel association models, ablation experiments were conducted on three datasets: ISRUC-S1, ISRUC-S3, and SHHS1, using a conventional multi-view feature fusion model and the present invention's CCAM method. In this approach, the present invention constructs a feature fusion-based multi-view model that learns and combines features from multiple views, particularly features from channels C3-A2. The experimental results, shown in Table 10, demonstrate that the present invention's CCAM method has a significant advantage over conventional multi-view models in enhancing classification performance based on single-channel data. This highlights the advantages of the cross-channel correlation model.

[0110] Table 8

[0111]

[0112] Table 9

[0113]

[0114] Table 10

[0115]

[0116]

[0117] To enable more interpretable analysis, this invention utilizes t-SNE to visualize feature embeddings based on single-channel FeatureNet, the Multi-view feature fusion method, and the CCAM method of this invention. Results on the ISRUC-S3 dataset are shown below. Figure 5 As shown. For use Figure 5For FeatureNet with single-channel data in (a), different categories are generally distinguishable, but the interfaces between categories are unclear. In particular, the feature embeddings of sleep stage N1 are severely confused with those of other stage categories. N1 sleep is also difficult to identify due to its short duration, shallow sleep, and indistinct features. Notably, even in this case, the feature embeddings of sleep stage categories N1 and N3 do not overlap. This is because there is a significant difference between deep and light sleep, and the features of sleep stages do not jump from N1 to N3. The visualization results of this invention are consistent with clinical findings. Figure 5 (b) In traditional multi-view feature fusion, the boundaries between categories are clearer and there is less overlap. In contrast, the CCAM method of this invention... Figure 5 As shown in (c), the category boundaries are very clear, and the overall classification error is small. Each category clusters together, especially the wakefulness, N1, and N3 sleep stage categories, and the clustering of each category tends to be cyclical.

[0118] To enable more interpretable analysis, this invention utilizes t-SNE to visualize feature embeddings based on single-channel FeatureNet, the Multi-view feature fusion method, and the CCAM method of this invention. Results on the ISRUC-S3 dataset are shown below. Figure 5 As shown. For use Figure 5 For FeatureNet with single-channel data in (a), different categories are generally distinguishable, but the interfaces between categories are unclear. In particular, the feature embeddings of sleep stage N1 are severely confused with those of other stage categories. N1 sleep is also difficult to identify due to its short duration, shallow sleep, and indistinct features. Notably, even in this case, the feature embeddings of sleep stage categories N1 and N3 do not overlap. This is because there is a significant difference between deep and light sleep, and the features of sleep stages do not jump from N1 to N3. The visualization results of this invention are consistent with clinical findings. Figure 5 (b) In traditional multi-view feature fusion, the boundaries between categories are clearer and there is less overlap. In contrast, the CCAM method of this invention... Figure 5 As shown in (c), the category boundaries are very clear, and the overall classification error is small. Each category clusters together, especially the wakefulness, N1, and N3 sleep stage categories, and the clustering of each category tends to be cyclical.

[0119] Compared with existing technologies, the proposed solution of this invention is a sleep stage classification method for universal sleep monitoring scenarios. The system mainly includes three modules: effective representation of single-channel physiological signals, pairwise correlation mining between channels, and multi-view sleep stage classification. The effective representation of single-channel physiological signals is used to extract and fuse multi-scale information of sleep physiological signals, and serves as the backbone model of a naive single-channel feature extraction and classification model as well as a cross-channel correlation network. The pairwise correlation mining model between channels is driven by the fusion of data and intelligent algorithms. It utilizes multi-channel data and, in conjunction with pairwise cross-channel correlation mining algorithms, transfers multi-channel PSG signal information to single-channel EEG signals, significantly enhancing the sleep stage classification effect in a universal environment based on single-channel EEG. Through multi-view fusion, only a single EEG channel is needed in the inference stage to fuse all the information of multi-channel PSG from multiple angles, thereby achieving a more effective and universal single-channel sleep stage classification.

[0120] 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.

[0121] 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.

[0122] 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.

[0123] 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 method for constructing a single-channel EEG sleep classification model, characterized in that, The method includes: S1. Obtain 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, select the target single channel data and the original associated channel data from each record in the original PSG data to form a training dataset; wherein, the other multiple channels are multiple channels among EEG channels, ECG channels, EMG channels, and EEG channels other than the target single channel; S3. Construct an initial classification model, which includes an initial association feature mining unit, a classification unit, and a fusion module, wherein: The initial association feature mining unit is used to extract features of target single-channel data and mine features of multiple association channels based on the features of target single-channel data. It includes a target single-channel feature extraction module and multiple initial association feature mining modules. Each initial association feature mining module includes an original association channel data feature extraction module and an association feature mining module. The classification unit is used to perform sleep stage classification prediction based on the target single-channel features and the features of multiple associated channels mined from 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. Train the initial classification model using the training dataset until convergence to obtain a model consisting of a target single-channel feature extraction module, an association feature mining module, a classification unit, and a fusion module. During training, the model parameters are updated using a comprehensive loss constructed from the distance loss between the association channel features mined from the target single-channel data and the features of the association channel data, and the classification label loss. The distance loss is: in, Indicates the first The maximum average difference loss between the features of the original associated channel data mined by the initial associated feature mining module and the features of the associated channels mined by the associated feature mining module based on the target single-channel features in each initial associated feature mining module. It is the first The weights of the loss of each initial association mining module, and in This represents a kernel function that maps data to a high-dimensional space. Indicates the target single channel number Features of each data sample Indicates the first The initial association mining module in the first association mining module is based on the target single channel. Features of associated channels mined from a data sample Indicates the first The initial association mining module in the first association mining module is based on the target single channel. The data sample corresponding to the first The characteristics of the associated channels were mined from the original associated channel data.

2. The method according to claim 1, characterized in that, The target single-channel feature extraction module is configured as the FeatureNet model.

3. The method according to claim 1, characterized in that, Each original associated channel data feature extraction module is used to extract the features of its corresponding original associated channel data, and each associated feature mining module is used to mine the features of its corresponding associated channel based on the features of the target single channel data.

4. The method according to claim 3, characterized in that, The classification unit includes multiple classification modules, one of which is used to classify sleep stages based on the target single-channel features or the features of the mined associated channels.

5. The method according to claim 4, characterized in that, The preset EEG channel is either the C3-A2 channel or the C4-A1 channel.

6. The method according to claim 5, characterized in that, The other multiple channels are two EEG channels, one ECG channel, two EOG channels, and one EMG channel, which are not included in the target single channel.

7. The method according to any one of claims 3-6, characterized in that, The overall loss is: in, Indicates the loss of classification labels. It is a distance loss, and: in, This represents the data sample size of the target single-channel data. Indicates the number of initial association mining modules. This represents the total number of original association channel data feature extraction modules and association feature mining modules corresponding to all initial association mining modules in the association mining unit. Represents cross-entropy loss. Indicates the target single channel number The data sample at the ... Features obtained from each module Indicates the target single channel number The data sample at the ... The predicted classification label is obtained by classifying and predicting the features acquired by each module. Indicates the target single channel number The actual classification label corresponding to each data sample.

8. A single-channel EEG data sleep classification model based on the method of any one of claims 1-7, characterized in that, The model includes: The associated feature mining unit includes a target single-channel feature extraction module and multiple associated feature mining modules. The target single-channel feature extraction module is used to extract features of the target single-channel EEG data, and each associated feature mining module is used to mine features of an associated channel based on the features of the target single-channel EEG data. The classification unit includes multiple classification modules. Each classification module corresponds to a target single-channel feature extraction module or an associated feature mining module. Each classification module is used to perform sleep stage classification prediction based on the features obtained by its corresponding module to obtain multiple sleep stage classification prediction results. Each sleep stage classification prediction result is the probability that the current single-channel EEG data belongs to that sleep stage. The fusion module is used to fuse the sleep classification prediction results of each classification module to obtain the final sleep classification result of the target single-channel data, wherein the sleep classification prediction result with the highest probability is taken as the final sleep stage classification result.

9. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the electronic device to perform the steps of the method as described in any one of claims 1 to 7.

Images