Introduction

In 2007, the American Academy of Sleep Medicine (AASM) developed new sleep interpretation guidelines - The AASM Guidelines. The AASM guidelines define sleep stages as Wake, non-rapid eye movement (NREM), and rapid eye movement (REM). NREM is divided into NREMI stage (N1), NREMII stage (N2) and NREMIII stage (N3). When falling asleep, the sleep stage of normal people first enters the NREM stage, gradually shifts from N1 stage to N3 stage, and then enters the REM stage at N2 stage or N3 stage. The sleep stage passes from NREM to REM, representing a complete sleep cycle^1–5.

Sleep staging results can be linked to certain diseases. Rapid eye movement sleep behavior disorder (RBD), as an early biological marker for predicting neurodegenerative diseases, such as Parkinson’s disease, is of great significance for the prevention and treatment of the latent stage of neurodegenerative diseases. RBD diagnosis requires accurate REM staging.² Depressed patients had low sleep efficiency and poor sleep quality. Accurate and detailed sleep staging results are required for both diagnosis and pharmacological intervention, particularly N3.^3,4 The proportion of REM sleep time is reduced in adolescents with bipolar depressive disorder (BDD). The duration of N2 was shorter in adult BDD patients than in normal adults.⁵ Although the diagnosis of sleep apnea syndrome (SAS) requires respiratory indications, and blood oxygen indications, the results of sleep staging also play an important role in the determination and research of SAS of the central and obstructive types.⁶

Clinically, polysomnography (PSG) is generally used for sleep-stage classification in patients.⁷ It takes a doctor two hours to determine a patient’s sleep stage using PSG (particularly electroencephalogram (EEG)). The workload is large and it depends on the experience of the doctor. Patients are required to wear EEG electrodes overnight. The measurement method is inconvenient and subject to greater subjective influence.^8,9 The lack of medical resources and low diagnostic efficiency have resulted in some shortcomings in professional diagnosis of related diseases and insufficient popularization of sleep quality preliminary screening. Therefore, an efficient sleep staging diagnostic method is needed to assist clinical decision making.

Many researchers attempt to develop automatic sleep staging methods. The main signal of the automatic sleep staging algorithm is EEG. Since the 1980s, artificial intelligence has been rapidly developing and has become a research hot spot. There have been numerous studies on the combination of machine learning and sleep staging in the past decade. In 2010, Güneş et al used a novel algorithm known as k-means clustering-based feature weighting (KMCFW) to increase the accuracy from 55% to 82%.¹⁰ A study by Silveira in 2017 had an accuracy higher than 90%. The method consisted of decomposing EEGs using a discrete wavelet transform and computing the kurtosis, skewness, and variance of its coefficients at selected levels. A random forest predictor was trained to classify each epoch into one of the R or K’ stages.¹¹ Seifpour designed an accurate and robust computer-assisted sleep stage scoring system using a single-channel EEG signal by proposing a novel time-domain feature known as statistical behavior of local extrema (SBLE), which extracted and defined various patterns in 2018. The average accuracy rate was 97.9% for the six-stage to two-stage sleep classification of the Sleep-EDF dataset.¹² Fifty-five time and frequency-domain features were extracted from the EEG signals and were then constituted as the inputs to the LSTM networks. The average accuracy rate was 83.6% on the Sleep-EDF dataset.¹³ Fang and Tao of South China Normal University achieved an accuracy of 88% based on the energy characteristics of EEG signals and the regular changes in fuzzy entropy in different sleep stages, combined with support vector machine classification.¹⁴ Wenbing et al extracted the relative energy mean of the rhythm wave using the wavelet transform algorithm and the multi-scale entropy of 9 to 13 and input it into the feedback propagation neural network classifier. Through the statistical analysis of the experimental results, the average staging accuracy of the proposed method was obtained as 85.81%.¹⁵

In this study, 120 training sets and 80 test sets were randomly composed of the first 200 groups of data from the National Sleep Research Resource Database of the United States. Two types of classifiers were used in the database. Then, the superior performer was selected, optimized, and used on the clinical data. The accuracy, precision, recall, and F1 score were used to evaluate the algorithm.

Materials

Shhs1

The SHHS1 database¹⁶ contains considerable sleep data of healthy subjects. The Sleep Heart Health Study (SHHS) conducted by the National Heart, Lung, and Blood Institute shares the physiological signal data collected during the clinical sleep trials in the National Sleep Research Repository. Among them, the PSG experimental samples in the SHHS1 dataset were derived from 6441 individuals collected from 1995 to 1998.¹⁷ The subjects were all aged 40 years or older, had no sleep-disordered breathing syndrome, and had not undergone tracheostomy and home oxygen therapy. Each sample contains C3-A2 and C4-A1 EEG, left and right eye, EMG, and ECG signals, and each sample contains experts’ sleep staging results in a 30-second frame. There are six movement categories. In this study, the first 200 data groups in the SHHS1 dataset were indiscriminately selected as experimental data, and the staging results annotated by experts were used as the standard staging results. The EEG adopts the third lead C3-A2 signal, the EMG adopts the fifth lead EMG signal, and the EOG adopts the sixth lead EOGL signal.

Clinical Data

This study complies with the provisions of the Helsinki Declaration and has obtained the written informed consent of the patients. The study was approved by the Human Research Ethics Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine. The batch number is (2021) LunShenYanDi (0053). The PSG instrument was used to collect 51 real individuals, aged 18–65 years, without sleep respiratory syndrome and without tracheostomy or home oxygen therapy. Each measurement should take no less than 8 hours. The subjects had no sleep-related diseases and had a good schedule. In the experiment, the environment should be quiet, no noise, no light stimulation and the temperature should be moderate. C3 EEG, left and right eye EEG, and EMG were collected, and the sleep results were recorded at one frame every 10 seconds. The results were calculated at one frame in 30 seconds by integrating algorithms, and the staging results were evaluated by doctors. Finally, 20 data that meet the requirements were sorted out.

Methods

The algorithm of this project included signal preconditioning, feature extraction and screening, classifier design, stage adjustment, and other steps.

Signal Processing

Preconditioning

ECG, EEG, EMG, and EOG signals are weak physiological signals with low amplitude and are easily disturbed. The noise was first removed. The EMG and EOG signals have distinct time-domain characteristics, so the baseline and poles were first removed; then, a zero-stage filter was used for filtering. The effective frequency band of eye electricity was mainly concentrated in the 0–10 hz range, and the passband frequency of the EOG filter was selected as 0.5–10 hz. The effective frequency band of EMG was 2–500 hz, and the passband frequency of the EMG filter was selected as 5–50 hz. The passband frequency of the EEG filter was selected as 0.3–50 Hz.¹⁸

Feature Extraction and Screening

The EMG amplitude was the highest in the wake stage and the lowest in the REM stage. The amplitude of EOG was larger in the wake stage and REM stage. Therefore, the relationship between EOG, EMG and sleep phases is mainly reflected in the time domain. Different sleep stages have different brain electrical rhythms, including α rhythm, slow wave activity during sleep, apex wave, sleep spindle wave and K complex wave. The difference of EEG rhythm is mainly reflected in the frequency domain, so the relationship between EEG and sleep stages is mainly reflected in the frequency domain.

For the non-stationary signal with weak amplitude, there are many time domain characteristics, such as integral mean, variance and coefficient of variation. Among them, the mean can express the signal strength, the variance can express the signal activity, and the coefficient of variation can balance the mean and variance and unify the dimension. In the wake stage and REM stage, the integral mean value and variance amplitude of EOG are significantly larger than those in other sleep stages, which is consistent with the physiological characteristics of the more active EOG in the wake stage and REM stage mentioned above. Therefore, the time-domain characteristics proposed above are of great significance for distinguishing the wake stage, REM stage and other sleep stages. The variance of EMG is the largest in the wake stage, and with the deepening of sleep, the characteristic value gradually decreases and approaches zero in the rapid eye movement stage, which accords with the weakest characteristic of EMG in the rapid eye movement stage analyzed above. Wake and REM stages can be distinguished according to the amplitude of the EMG features. According to the AASM standard, artificial staging mainly depends on the distribution of different EEG rhythms to judge the sleep stage, and EEG rhythms are signals located in different frequency bands. Therefore, frequency domain features of EEG signals can be extracted to analyze the dominant EEG rhythms in different stages, so as to determine the sleep stage according to the distribution of EEG rhythms. In this paper, spectral relative energy, spectral energy ratio of different EEG rhythms and relative power of different frequency bands are used to describe the frequency domain of EEG signals.

The quantity of the feature parameters and the correlation between the feature parameters and classification results directly affect the classification results. It is required to select several features that contribute the most to the classification. First, the features are standardized. Then, the T-test, variance filtering, mutual information, embedding, recursive feature elimination, and other methods were used to select the features. For feature dimension reduction, filtering method is used to remove some irrelevant features, embedding method is used to rank the importance of each feature, packaging method is used to select the least and best feature collection under the premise of no decline in accuracy, and PCA method is used for dimension reduction. Finally, 20 features are selected. The stratified sampling K-fold cross-validation method was used to verify the performance of the classifier. Thus, simplified features with a large amount of information and weak cross-correlation were obtained.

Support Vector Machine (SVM) Algorithm and Extreme Gradient Algorithm

Svm

The core purpose of the SVM classifier is to separate two or more samples by constructing a split surface to determine the separating hyperplane, which correctly divides the training dataset and has the largest geometric separation. SVM is suitable for solving small-sample, high-dimensional, and nonlinear problems, and can improve generalization performance.^19,20

First, the eigenvalues were uniformized to improve the model accuracy and convergence speed. Then, a kernel function and different parameters were selected to set the model. Different parameters affect the correctness and accuracy of classification and tolerance to errors. Finally, several parameters were selected for testing to determine the optimal parameters.

When using the SVM algorithm, the data was manually divided into segments of 10s for algorithm processing. If a staging result violates the normal sleep physiological situation, it can be corrected by referring to “Clinical Electroencephalography”.^21,22

Extreme Gradient Boosting Algorithm

The extreme gradient boosting algorithm (XG Boost) is a variant of the gradient-boosting machine. Its purpose is to breakthrough the computing power limit of the gradient boosting tree algorithm and improve the performance and computing speed of the model. The algorithm was easy to parallelize, had high accuracy, and supported regression and classification. Additionally, it is not sensitive to the dimension, unit of the feature, and missing values; thus, it is not required to standardize the feature.^23,24

The XG Boost algorithm also requires setting and adjusting parameters. This project aims to use the grid search and K-fold cross-validation methods to optimize the parameters. Grid search is used to traverse all the parameters to determine the best groups: num_round = 300, eta = 0.275, max_depth = 6, γ=15, colsample=1, λ = 10, α = 0. The K-fold cross-validation method is used to train and test the complementary subsets, which were obtained by dividing the sample set multiple times. Its purpose was to avoid the overfitting caused by reusing a part of the data in the training model.^25,26

The results were also required to be logically adjusted according to the physiological situation and regularities of the sleep cycle.

The convergence speed of the SVM and the accuracy of the model is closely connected with standardization. If the data is not normalized, it will take several hours to run the algorithm. Therefore, eigenvalue standardization must be performed for the support vector machine. On the contrary, for XG Boost, if the features are standardized, some data information will be lost, which affects the model accuracy.

SVM and XG Boost were applied to the database. Then, 40 sets of data were selected indiscriminately in SHHI. As shown in Table 1, the average accuracy of the XG Boost algorithm is 76.77%, and the average accuracy of the SVM algorithm is 63.37%. The SVM algorithm requires approximately 20s to predict a set of data, whereas the XG Boost algorithm requires less than 1 s. In terms of accuracy and time, XG Boost was superior.

Table 1 The Accuracy of the First 40 Test Sets on Different Algorithms

Deeper Insights from Classification Models

Algorithm Optimization

For the XG Boost algorithm, the generalization error of the model can generally be evaluated and adjusted through experimental tests. For the XG Boost algorithm, the generalization error of the model can generally be evaluated and adjusted through experimental tests. Owing to the problem of sample imbalance in sleep staging, stratified sampling was used to ensure that the proportion of sample categories was balanced. This study used the 5-fold cross-validation method.

Sample from SHHS1 sizes varies with sleep stage. The Wake stage accounted for 31.59%, N1 stage accounted for 3.24%, N2 stage accounted for 39.69%, N3 stage accounted for 13.27%, and REM stage accounted for 12.21%.

If the problem of sample imbalance is ignored, poor specificity and sensitivity of this category, which contains fewer samples, occur. XG Boost algorithm can set the weight of each sample to solve the problem of sample imbalance. First, the weight of each category is directly set as the reciprocal of the proportion of samples in this category, so that the weight of samples in each category can be regarded as equal. Then, an exhaustive test was carried out near this set of parameter combinations, and the weights of N1, N3 and REM stages were traversed with a certain step length, and the average accuracy of stratified sampling 5-fold cross-validation was used as an evaluation index to select the parameter model with the best score. The weight of wake is finally determined to be 3.2, N1 is 16.0, N2 is 2.5, N3 is 7.0, and REM is 8.0. After the weight setting, the recall rate of stage N1 increased 23%, and the F1 score has increased 20%.

Correction of Sleep Staging Results

The staging results obtained using the XG Boost algorithm are only the results of machine learning, often violating the physiological characteristics. Therefore, three characteristics were followed through the regularity of the sleep cycle to revise the results of sleep staging in this study:

The initial state of sleep data is generally the Wake stage, and the duration will be no less than 1 min;

The REM stage is generally transformed from the N2 and N3 stages, and then, the REM stage is generally transformed into the N2 and N3 stages;

Generally, from the N1 and N2 stages of the two light sleep stages instead of the N3 and REM stages into the Wake stage.

Conclusion and Discussion

Table 2 is the confusion matrix. The sum of each row of the confusion matrix is the actual number of samples of that category, and the sum of each column is the predicted number of samples of that category. Nij is the number of samples that are actually category i but predicted to be category j. i and j are 0, 1, 2, 3, 5, which correspond to the five sleep stages of Wake, N1, N2, N3, and REM, respectively.

Table 2 Confusion Matrix

This study used precision, recall, accuracy, and F1 score to comprehensively evaluate the model.

Tables 3–6 lists the result that the algorithm applied to SHHS1 test samples. Tables 7–11 list the result that the algorithm applied to clinical data.

Table 3 The Precision of 80 Sets of Test Data

Table 4 The Accuracy of 80 Sets of Test Data

Table 5 The F1-Score of 80 Sets of Test Data

Table 6 The Recall of 80 Sets of Test Data

Table 7 The Precision of 20 Clinical Data

Table 8 The Accuracy of 20 Clinical Data

Table 9 The F1-Score of 20 Clinical Data

Table 10 The Recall of 20 Clinical Data

Table 11 The AUC of 20 Clinical Data

As shown in Tables 3–6, in the database,

1. The average accuracy of 80 groups of samples is 83.24%.
2. For the Wake and N2 stages, the precision and recall rates are both over 80%. The algorithm has the strongest staging ability in the Wake and N2 stages.
3. The algorithm has the second highest staging ability in N3 and REM. For the N3 and REM stages, the precision and recall rates are above 70%. Few samples were misjudged and missed.
4. For the N1 stage, the algorithm staging ability is low.

In the clinical data,

1. The average accuracy of 20 groups of samples is 76.37%.
2. For the Wake and N3 stages, the precision and recall rates are both above 85%. The algorithm has the strongest staging ability in the Wake and N3 stages.
3. The algorithm has the second strongest staging ability in N2 and REM. For the Wake, N2, and REM stages, the recall rates are over 70%. Few samples were misjudged and missed.
4. For the N1 stage, the algorithm staging ability is low.
5. The AUC value of each sleep stage is 0.7637.

The sleep staging result of the algorithms, which run in the database and clinical data, are similar. The wake stage exhibits the best staging results because its brain wave characteristics are significantly different from those of the other stages. The N1 stage is the transition stage between the Wake and N2 stages. It is easy to misjudge the N1 stage as the Wake and N2 stages, so the ability to recognize the N1 stage is unfavorable. The main EEG of the N2 and N3 stages is sleep spindle and K wave. The characteristics of the two stages are similar, so the staging results of N2 and N3 are similar. In clinical trials, there may be more objective factors interfering with the REM stage, so the accuracy of REM stage recognition is weak.

Using a computer to automatically identify sleep stages can reduce the workload of doctors and achieve the effect of assisting diagnosis.

We compare XG Boost in this paper with other 2 traditional machine learning models (Support Vector Machine and Random Forest) on publicly available datasets as shown in Table 12.²⁷ Among these machine learning methods, XG Boost algorithm performs better than SVM and RF.

Table 12 Performance Comparison of Different Methods (F1-Score)

We compare XG Boost in this paper with deep learning model GAC-SleepNet in Table 13. GAC-SleepNet uses the characteristic information in the dual structure of the graph structure and the Euclidean structure for the classification of sleep stages. In the graph structure, this study uses a graph convolutional neural network to learn the deep features of each sleep stage and converts the features in the topological structure into feature vectors by a multi-layer perceptron. In the Euclidean structure, this study uses convolutional neural networks to learn the temporal features of sleep information and combine attention mechanism to portray the connection between different sleep periods and EEG signals while enhancing the description of global features to avoid local optima.²⁸ For stages with less obvious features, the accuracy of the GAC-Sleep NET algorithm is higher than that of the XB Boost. This shows that the results of machine learning are more dependent on features. Although some deep learning models can achieve high precision, GAC-Sleep Net models are difficult to tune and optimize. So both kinds of methods have advantages and disadvantages. However, multi-modal deep learning is the trend of development so how to combine the advantages of machine learning and deep learning is worth studying.

Table 13 Performance Comparison of Different Methods (Precision)

Prospects of Interpretability Methods

The algorithm proposed in this study can be further improved and explored:

1. Optimized feature extraction. In this study, we extracted the time domain, frequency domain, and nonlinear characteristics of the signal and attempted to improve the accuracy of the model by extracting more relevant features.
2. Feature dimension reduction. In this study, the PCA method was used to select the optimal feature of 20 dimensions. However, it reduced the accuracy of the model. Other methods for feature dimensionality reduction can be attempted to ensure model accuracy.
3. The imbalanced samples. This study increased the weight of minority class samples such as N1. However, the precision and recall rates of the N1 stage are both low. Deep exploration can be performed with the physiological characteristics of N1.
4. More clinical trials. More clinical trials can be conducted to verify the stability of the system and the clinical applicability of the algorithm.
5. Explore deep learning models and continuously optimize the models.