An Ensemble Learning Sleep Staging Method Based on Clustering Dimensionality Reduction

Abstract

An ensemble learning sleep staging method based on clustering dimensionality reduction includes the following steps: acquiring single-channel EEG data and simultaneously initializing an SVM classifier; extracting features from the EEG data, extracting a feature vector for each time period as a data point; performing feature selection and feature classification, selecting feature sets for the classifier and training the base classifier; for test data, selecting the stage with the highest probability as the prediction result for the test data; this invention uses single-channel EEG signals as input, greatly reducing the difficulty of acquiring input signals, and the proposed feature selection and feature classification methods have good adaptability to different feature spaces; therefore, this invention can perform sleep staging quickly, accurately, and in real time.

Description

Technical Field

[0001] This invention belongs to the field of medical informatics technology, and specifically relates to an ensemble learning sleep staging method based on clustering dimensionality reduction, which uses SVM (Support Vector Machine) as the base classifier to achieve sleep staging. Background Technology

[0002] Sleep occupies one-third of a person's life, and its quality is closely related to physical and mental health. Related research shows that disruption of normal sleep patterns can lead to psychological and neurological imbalances. Furthermore, sleep disorders can have a more profound impact on cognitive function, such as decreased attention, reduced learning ability, and memory decline. Sleep staging refers to dividing the sleep process into different stages based on the periodic changes in physiological indicators such as electroencephalogram (EEG), electrooculogram (EOG), electrocardiogram (ECG), and blood oxygen saturation during sleep. The diagnostic results of sleep staging are not only important for the diagnosis of sleep disorders but also a prerequisite for neurointerventional treatment during sleep. [2] The AASM standard (American Academy of Sleep Medicine) is commonly used in current sleep EEG research. This standard divides sleep into the wake stage, non-rapid eye movement (NREM) sleep, and rapid eye movement (REM) sleep. The NREM stage is further subdivided into three periods: N1, N2, and N3.

[0003] However, traditional sleep staging is determined by doctors through observation of physiological signal data from polysomnography (PSG). Relying on doctors to perform sleep staging based on observation of entire night's sleep data is extremely labor-intensive, difficult to analyze in real time, and the accuracy of sleep staging results heavily depends on the doctor's experience. Therefore, the field of sleep medicine urgently needs an automated sleep staging method that can accurately perform sleep staging.

[0004] Existing sleep staging algorithms are divided into two types: multi-channel signal processing and single-channel signal processing. For example, the article on multi-channel algorithms, "S. Khaliighi, T. Sousa, G. Pires, U. Nunes, Automatic sleep staging: a computer assisted approach for optimal combination of features and polysomnographic channels, Expert Syst. Appl. 40(17)(2013)7046–7059," states that while adding more physiological signal channels can indeed improve the accuracy of the classification algorithm, it also incurs greater time and memory overhead. Moreover, multi-channel algorithms rely on more signal acquisition devices, which not only makes the diagnostic process more cumbersome, but also introduces random noise due to interference between multiple devices, making the algorithm performance unstable. Furthermore, if real-time sleep staging is to be achieved, multi-channel processing algorithms are difficult to compare with single-channel algorithms.

[0005] Numerous studies have demonstrated that single-channel EEG signals are sufficient for sleep staging, as illustrated in "Boostani R, Karimzadeh F, Nami MA comparative review on sleep stage classification methods in patients and healthy individuals[J]. Computer Methods and Programs in Biomedicine. 2016, 140(2017):77-91." Currently, sleep staging algorithms for single-channel EEG signals are mainly divided into two categories: traditional machine learning algorithms based on feature extraction and end-to-end deep learning algorithms. Research indicates that traditional machine learning methods can achieve classification performance comparable to deep learning algorithms in the field of sleep staging. The patent application CN202310356676.3, "Sleep Staging Analysis System and Method Based on Temporal Self-Attention and Dynamic Graph Convolution," accepts multi-channel EEG features as input and employs a self-attention mechanism and dynamic graph convolution as a classification model to achieve sleep staging. While this method achieves good classification results, the multi-channel EEG signal input makes the signal acquisition process more complex. Furthermore, without GPU acceleration, deploying attention networks and dynamic graph convolutions on the vast sample space of multi-channel EEG data inevitably incurs significant computational overhead. For sleep staging algorithms that need to be deployed on therapeutic devices, such deep learning networks are unsuitable. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, the present invention aims to provide an integrated learning sleep staging method based on clustering dimensionality reduction. Considering that the labels of the training data are manually labeled by physicians and there is a certain probability of bias, a clustering-based feature selection method is proposed. SVM is used as the base classifier, and the algorithm has low time and memory overhead. In addition, using a single-channel EEG signal as input greatly reduces the difficulty of input signal acquisition. Furthermore, the proposed feature selection and feature classification have good adaptability to different feature spaces. Therefore, the present invention can perform sleep staging quickly, accurately, and in real time.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] An ensemble learning sleep staging method based on clustering dimensionality reduction includes the following steps:

[0009] Step 1: Obtain single-channel EEG data and sleep staging diagnosis results from professional physicians; perform bandpass filtering on the EEG data for each time period; and initialize the SVM classifier.

[0010] Step 2: Extract features from the EEG data, extracting one feature vector for each time period as a data point;

[0011] Step 3: Perform feature selection and feature classification. Select feature sets for the classifier, select the features to be retained, and assign the retained features to the feature sets.

[0012] Step four: Train the model. The sleep stage diagnosis results of the professional physicians in step one are used as labels for the training data. The feature subsets selected for each classifier in step three are used to generate five subspaces from the feature space, which are used to train the five base classifiers respectively. Each base classifier is trained independently. After training, each base classifier can give a prediction of a sample, which is the probability of the corresponding sleep stage being present or absent.

[0013] Step 5: For the test data, {SVM0, SVM1, SVM2, SVM3, SVM4} give the probability of whether the corresponding sleep stage is present or not. After integrating the outputs of the five classifiers, the probability of the test data being classified into five stages is obtained. The stage with the highest probability is selected as the prediction result of the test data.

[0014] Step one specifically involves: first, dividing the EEGG data into time periods, using each time period as the smallest processing unit, and applying a Butterworth filter for bandpass filtering from 0.5 to 45 Hz. Simultaneously, initializing five SVM classifiers, {SVM0, SVM1, SVM2, SVM3, SVM4}, as base classifiers for the classification tasks of the five sleep stages, with the SVM numbers corresponding to the sleep stages.

[0015] The implementation method of step three is as follows:

[0016] (a) Data augmentation of the feature space, i.e., using the method of copying existing data points to make the amount of data in each sleep stage in the feature space the same, resulting in an augmented overall dataset. in l i ∈{0,1,2,3,4},(p i ,l i ) represents a sample;

[0017] (b) Regarding feature x i Construct a dataset D(x) containing only this feature. i );

[0018] (c) Regarding D(x) i Perform k-means binary clustering to obtain two clusters, cluster1 and cluster2.

[0019] (d) Calculate the information entropy of the two clusters respectively;

[0020] (e) Select the cluster with low information entropy as the label cluster CC(x) for this feature. i ), statistics CC(x i The number of samples with the actual label k in PL(x) is used to select the label with the largest number of samples and assign it to PL(x). i Clustering algorithms for CC(x) i The predicted value of all sample labels in ) is PL(x i );

[0021] (f) Feature x i Assigned to Feature set, feature x i The score P(x) i ) is CC(x i The percentage of samples in the cluster whose true label is the same as the cluster prediction label PL(i);

[0022] (g) For each feature, repeat (b)-(f) until all features are classified and scored, and P(x) is obtained. i Select the features you want to keep.

[0023] The specific implementation method of step five is as follows:

[0024] Each base classifier provides the probability of a sample being classified as belonging to the sleep stage corresponding to that classifier, i.e., output [p0, p1]. By comprehensively considering the output results of each model, the final probability vector [p(0), p(1), p(2), p(3), p(4)] of the sample being classified as one of the five stages is obtained. The calculation method for this vector is as follows:

[0025] Compared with the prior art, the advantages of the present invention are:

[0026] (1) In step one of this invention, a single-channel EEG signal is used as input, which greatly reduces the difficulty of acquiring the input signal. In particular, the combination with the ensemble learning method and cluster-based feature selection further optimizes the ratio of effect, accuracy, memory overhead and time overhead.

[0027] (2) This invention adopts the ensemble learning method and uses SVM as the base learner to realize sleep staging. The algorithm has low time and memory overhead, and the method can reach the forefront level in terms of sleep staging accuracy.

[0028] (3) The cluster-based feature selection proposed in step three of this invention can perform feature selection and feature classification for different feature spaces. Therefore, this algorithm has good adaptability to different feature extraction schemes.

[0029] In summary, compared with existing technologies, this invention proposes a cluster-based feature selection and classification method, and uses SVM as a base learner to construct a sleep staging ensemble learning model, which can achieve real-time sleep staging with low memory and time overhead. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention.

[0031] Figure 2 This is a flowchart of the feature extraction process of the present invention.

[0032] Figure 3 This is the feature selection process for the present invention.

[0033] Figure 4 This is a schematic diagram of the model training process of the present invention.

[0034] Figure 5 This is a schematic diagram of the prediction process of the present invention. Detailed Implementation

[0035] The invention will be further described below with reference to the accompanying drawings.

[0036] Figure 1 The overall construction process of the model of this invention is demonstrated. In the following text, (p,l) represents a sample point (p is the attribute vector of the sample, l is the sample label), x represents a feature, S represents a feature set, and D(S) represents the dataset constructed from feature set S. This invention is specifically implemented according to the following steps:

[0037] Step 1: Obtain single-channel EEG data and sleep staging diagnosis results from a professional physician. Perform bandpass filtering on the EEG data for each time period and initialize the SVM classifier simultaneously.

[0038] In this method, sleep stages are mapped to numbers, and the mapping relationship is shown in Equation (1). A 30s EEG data point is considered as one time period, and each time period is the smallest sleep stage unit. A Butterworth filter is used to perform bandpass filtering of the EEG signal from 0.5 to 45 Hz. Simultaneously, five SVM classifiers are initialized for the classification tasks of the five sleep stages: {SVM0, SVM1, SVM2, SVM3, SVM4}, with the SVM number corresponding to the sleep stage.

[0039]

[0040] Step two: Extract features from the EEG data, extracting one feature vector for each time period as a data point.

[0041] This method is applicable to different feature extraction schemes and feature sets, and does not restrict the types of features. This study used a feature set of 32 features, with feature numbers shown in Table 1. The extraction process is described below. Figure 2 Features 0-5 are time-domain features, extracted directly from the time-domain signal. Features 6-13, 22, 23, and 24 are extracted based on the signal's power spectrum, obtained using the Welch periodogram method. The frequency range of the EEG feature waveforms is shown in Table 2, and the power of the feature waves uses relative power, calculated using the formula: [Formula omitted for brevity]. Here, SubBand represents the frequency band range of the feature wave, and P... xx This represents the power spectrum of the signal.

[0042]

[0043] This invention also extracts the AR coefficients (Autoregressive coefficients) of the signal. An 8th-order autoregressive model was used to fit the time-domain data, resulting in eight AR coefficients (ar1-ar8), each existing as an independent feature. Based on the AR coefficients, this invention extracts the Itakura Distance (ItaDis), which represents the distance between two points represented by the AR coefficient vector. The specific calculation method can be found in "Itakura Distance: AUseful Similarity Measure between EEG and EOG Signals in Computer-aided Classification of Sleep Stages". This invention extracts the Itakura Distance from each point to the center of the five sleep stages, thus five features are extracted (ItaDis1-ItaDis5). Multiscale permutation entropy and fuzzy entropy are nonlinear characteristics of signals. The calculation methods can be found in "Yao Wenpo, Liu Tiebing, Dai Jiafei, Wang Jun. Multiscale permutation entropy analysis of EEG signals [J]. Acta Physica Sinica. 2014, 63(7): 078704." and "Peiyu Weng, Keming Wei, Tian Chen, et al. Fuzzy Approximate Entropy of Extrema Based on Multiple Moving Averages as a Novel Approach in Obstructive Sleep Apnea Screening [J]. IEEE Journal of Translational Engineering in Health and Medicine. 2022, 10: 4901211.".

[0044]

[0045]

[0046] Table 1

[0047]

[0048] Table 2

[0049] Step 3: Perform feature selection and feature classification. Select feature sets for the classifier, select the features to be retained, and assign the retained features to the feature sets.

[0050] The feature selection method in this invention not only performs dimensionality reduction on the overall feature space but also assigns different features to different classifiers. Since a single SVM can only perform binary classification, this invention requires five parallel base classifiers {SVM0, SVM1, SVM2, SVM3, SVM4} to achieve five-class classification, with each SVM base classifier performing a binary classification task to determine whether a sleep phase has occurred. This invention proposes selecting feature sets for each of the five classifiers. The method is as follows: S is the overall feature set S = {x0, x1, ..., x...} 31}(x i (For feature numbered i in Table 1). The feature selection process is described in [link to feature selection process]. Figure 3 S represents the overall feature set, D(S) represents the dataset constructed from the feature set S, xi∈S, D(xi) represents the feature subspace constructed from the feature xi, CC (Characteristic Cluster) represents the label cluster, PL (Predicted label) represents the predicted label value of the sample within the label cluster, and dim(S) represents the dimension of the feature space.

[0051] The specific implementation method is as follows:

[0052] (1) Perform data augmentation on the dataset, that is, for samples with fewer classes, duplicate the data so that the number of data in each class is the same, and obtain the overall dataset. in l i ∈{0,1,2,3,4},(p i ,l i ) represents a sample (in this invention, the feature space dimension d = 32).

[0053] (2) Using feature x i Construct a dataset D(x) containing only this feature i ),

[0054] (3) Perform k-means binary clustering to obtain two clusters, cluster1 and cluster2, as shown in the equation.

[0055]

[0056] (4) For cluster1 and cluster2, calculate the information entropy of the clusters respectively, as shown in formula (4), where p k This represents the proportion of samples with the true label k within a cluster. The correspondence between label k and sleep stage is shown in the formula.

[0057]

[0058] (5) As shown in the equation, select the cluster with the smaller entropy as the label cluster CC(x) for this feature. i (Characteristic cluster, CC); Statistical CC(x) i The actual number of samples with label k in PL(x) is used to assign the label with the largest number of samples to PL(x). i (Predictedlabel, PL), the clustering model will use CC(x) i The predicted label for all samples within ) is PL(x i ), see formula.

[0059]

[0060] PL(x i ) = argmax 0≤k≤4 count(k) (6)

[0061] argmax: The independent variable used to maximize the objective function.

[0062] count(k): Counts the number of samples with label k.

[0063] (6) Feature x i Incorporation into feature set See the formula. The feature score P(x) i ) represents the proportion of samples in CC(i) whose true labels are the same as the cluster prediction labels PL(i), as shown in the following formula.

[0064]

[0065]

[0066] count(): Counts the number of samples that meet the given conditions.

[0067] size(): Counts the total number of samples in the set.

[0068] (7) For each feature x i (In this description, 0 ≤ i ≤ 31), repeat steps (2)-(6) until all features are classified and scored. Through P(x i To select the features to be retained, this invention uses 0.3 as the threshold for retaining features, i.e., P(x) i Features with a value ≥ 0.3 were retained, as shown in Table 3.

[0069]

[0070]

[0071] Table 3

[0072] Step 4: Model Training

[0073] The sleep staging diagnosis results from the professional physician in step one are used as labels for the training data.

[0074] Based on the results of the third step of clustering feature selection, this invention proposes a model that trains five different SVMs using different feature sets and then concatenates them in parallel for five-class classification. The model training process is described below. Figure 4 SVM1-SVM5 perform binary classification for Wake, N1, N2, N3, and Rem classes, respectively. Specifically, SVMi uses S... i The feature set (0≤i≤4) is used for training and testing, and the base SVM classifier used in this invention adopts the method proposed by Platt. [5] By training, this method can provide the probability of classifying the data into each category.

[0075] The fifth step, the model's prediction process for new data, is detailed below. Figure 5 Five SVMs respectively yielded probabilities p0 and p1 for classifying new data into the positive and negative classes of the corresponding sleep stages. The results of the five models were then superimposed to obtain the probabilities of the data being classified into the five sleep stages. The superposition method is shown in the equation, where p1(SVM) k ) represents the output of SVMk, p1 (the probability of being judged as a positive example), p0 (the probability of SVMk being positive). i Let p0 represent the probability of being a counterexample as output by SVMk. The final label prediction for each data point is the period with the highest probability (equation).

[0076]

[0077] sleep_stage = max i∈{0,1,2,3,4} pi) (10).

[0078] As can be seen from the above embodiments, compared with the prior art, the present invention proposes a clustering-based feature selection and classification method, and uses SVM as a base learner to build a sleep staging ensemble learning model, which can achieve real-time sleep staging with low memory and time overhead.

Claims

1. A cluster-based dimension reduction integrated learning sleep staging method, characterized in that, The method comprises the following steps: Step one, obtaining single-channel EEG data and sleep staging diagnosis results of professional physicians, performing band-pass filtering on each period of EEG data, and initializing SVM classifiers; Step two, performing feature extraction on the EEG data, extracting a feature vector for each period as a data point; Step three, performing feature selection and feature classification, selecting feature sets for the classifiers, selecting features to be retained, and assigning the retained features to the feature sets; Step four, training the model, using the sleep staging diagnosis results of professional physicians in step one as the labels of the training data; using the feature subsets selected in step three to generate five subspaces from the feature space for training of the five base classifiers; each base classifier is independently trained, and after training, each base classifier can give a prediction of a sample to obtain the probability of being and not being the corresponding sleep stage; Step five, for test data, {SVM0, SVM1, SVM2, SVM3, SVM4} give the probability of being and not being the corresponding sleep stage, integrate the outputs of the five classifiers to obtain the probability of the test data being classified into five stages, and select the stage with the maximum probability as the prediction result of the test data; The step one is specifically: first, divide the EEGG data by period, take the period as the minimum processing unit, perform band-pass filtering of 0.5-45Hz using a Butterworth filter, and initialize five SVM classifiers as base classifiers for the classification tasks of the five sleep stages, {SVM0, SVM1, SVM2, SVM3, SVM4}, and the SVM number corresponds to the sleep staging; The implementation method of step three is: (a) data augmentation on the feature space, i.e. using the method of replicating existing data points so that the data amount of each sleep stage in the feature space is the same, obtaining the augmented overall data set wherein l i ∈{0,1,2,3,4},(p i ,l i ) represents a sample; (b) selecting a feature x i , constructing a dataset D(x i ) containing only this feature. (c) performing k-means biclustering on D(x i ) to obtain two clusters cluster1, cluster2; (d) respectively calculating the information entropy of the two clusters; (e) Select the cluster with low information entropy as the label cluster CC(x) for this feature. i ), statistics CC(x i The number of samples with the actual label k in PL(x) is used to select the label with the largest number of samples and assign it to PL(x). i Clustering algorithms for CC(x) i The predicted value of all sample labels in ) is PL(x i ); (f) assigning features x i to a feature set, the score P(x i ) of a feature x i is the proportion of samples in CC(x i ) for which the true label is the same as the cluster predicted label PL(i). (g) for each feature, repeating (b)-(f) until all features are classified and scored by P(x i ) selecting features that need to be retained; The specific implementation method of step five is: Each base classifier gives the probability that the sample is judged as the sleep stage corresponding to the classifier, that is, the output [p0, p1], and by comprehensively considering the output results of each model, the final probability vector [p(0), p(1), p(2), p(3), p(4)] of the sample being judged as the five stages is obtained, and the calculation method of the vector is

Images