1 Introduction

Each and individual human needs to have proper sleep at night and it is the basic requirement for the human being and proper sleep quality plays a direct role in our day-to-day life. Sometimes its impact is reflected in our physiological activities such as the quality of learning ability, physical activity, mental ability, and performance of the overall activities (Aboalayon et al. 2016). In general complete sleep, duration is covering through sleep stages at regular interval time and which is related to our brain neuron system (Chung et al. 2009). With the modern digital generation, the lifestyle of the human being is complicated and ultimately it has resulted that millions of people get a poor quality of sleep during night time. This problem is seen across the world with all age groups of people, and this is a global challenge in the health care sector because it has found from the different research that poor quality of sleep is the major responsibility of creating critical diseases such as bruxism (Heyat et al. 2019), insomnia (Lai et al. 2019), narcolepsy (Rahman et al. 2016), obstructive sleep apnea (Kim et al. 2018) and rapid eye movement behavioral disorder (Siddiqui et al. 2016). Finally, it creates damages to various parts of our body such as heart failure, brain stroke, and several neurological disorders. Normally, sleep behavior is characterized by changes in activities in our body such as respiration and heartbeat rate, brain-behavior, and muscle movements. Currently, there are two important sleep standards followed during sleep staging analysis. According to both standards the whole sleep stages are divided into three basic categories: (1) wakefulness (W), (2) rapid eye movement (REM), (3) non-REM. The first sleep handbook was edited by Rechtschaffen and kales (R&K) in 1968 (Rechtschaffen 1968) and under this sleep scoring rule, the NREM sleep stage is further divided into four sub-sleep stages: S1, S2, S3, and S4. These sleep standards followed all clinicians during the analysis of sleep irregularities of the patients, but in the year 2008, another recognized sleep recommendation was proposed by the American Academy of Sleep Medicine (AASM) (Iber 2007) and defined a new sleep handbook through small modification on the R&K rules. As per AASM rules, the non-rapid eye movement is further segmented into three sub-sleep stages such as N-REM1, N-REM2, and N-REM3 (Iber 2007). The sleep cycle generally repeated at regular intervals of time between NREM stages to REM stages and each duration of the sleep cycle is around 90–110 min (Carskadon and Dement 2017). The quality of sleep ratio is individually different from person to person according to their age.Stage1 of N-REM sleep (N-REM1) is light, where the subject's eye movements are slow and the muscle movements slow, in the N-REM2 sleep stage, the eye movements completely stop and the response of the brain becomes slower. So both N-REM1 and N-REM2 sleep stages are categorized as light sleep stages. Similarly, the N-REM3 and N-REM4 are called deep sleep, in which no eye movements have occurred and some muscle movements are appeared (Nagabushanam et al. 2019). Finally, in the REM stage, breathing is increased and the physical movements have been seen rapidly. Sleep staging is normally examined using polysomnographic (PSG) recordings from the admitted subject in the clinic.

Basically during the PSG test, different physiological recordings are collected from the subjects to measure the sleep quality during nighttime. The major recordings included during the assessment of sleep quality are electroencephalogram (EEG), electromyogram (EMG), electrooculogram (EOG), and electrocardiogram (ECG) (Holland et al. 1974). Among these, major researchers the priority toward an analysis of sleep pattern abnormality through EEG signal because the EEG signal provides the information on brain activities and behavior of subjects during sleep. It can help to characterize the behavior of whole sleep stages through different frequency ranges of EEG signal which were segmented into \(\delta \) band (< 3.5 Hz), \(\theta \) band (4–7 Hz), \(\alpha \) (8–13 Hz), \(\beta \) (14–30 Hz) and \(\gamma \) (30–80 Hz), ultimately it can support during the demonstration of the characteristics of different stages of sleep (Acharya et al. 2015). To correctly predict the sleep behaviors of the subjects, most of the clinicians used electroencephalogram, which directly provides information about brain activities. EEG recordings, its hectic situation for sleep experts to monitoring within the 30 s framework and fix the labeling of sleep stages (O’Reilly and Nielsen 2014; Hassan and Bhuiyan 2016a). This approach consumed more time and required more manpower for hours of sleep recordings. To overcome the difficulty of the manual approach, nowadays automated sleep scoring system is obtained to analyze the recorded EEG signals (Lei et al. 2018; Sharma et al. 2017; Zhu et al. 2014; Ronzhina et al. 2012) with subject to treatments various types of sleep and neurological disorders in real-time diagnosis. Some of these sleep scoring systems are based on polysomnographic signals recordings (Sharma et al. 2019a; Spriggs 2014; Hassan and Bhuiyan 2016b).

1.1 Literature review

The main aims of the literature study are analysis the existing contributions with subject to different methodologies and models used in recent sleep staging studies. Various computational methodologies were proposed by researchers to support sleep experts for assisting sleep staging. Those carried steps were on the information extraction (Polysomnography channel selection), on the preprocessing (removing the data artifacts and data normalization), on the feature extraction step (transformation of linear and nonlinear features), on the feature selection technique (identifying the most relevant features) and finally on the classification algorithm. These entire literature reviews are divided into two sections. In the first section, we discussed the different types of features extracted, and in the second part, we briefly discuss the different classification models used during sleep staging, which used recently sleep staging studies proposed for characterization the sleep-related abnormalities under sleep standards recommended by R&K and AASM manuals. Here we have presented some comparative studies regarding sleep staging.


Different Signal segmentation and feature extraction techniques used during Sleep staging

Zhu et al. (2014) proposed graph-based features for the analysis of the sleep patterns from EEG signals. The required sleep recordings were obtained from eight subjects having 14,963 epochs and the proposed SVM model achieved 87.5% classification accuracy.

Hassan and Bhuiyan (2016b) obtained wavelet transform concepts for decomposing the signal into different signal sub-bands using tunable Q-factor and extracted the spectral features. The entire experimental work conducted using 15,188 epochs from 28 subjects.

In Hassan and Bhuiyan (2016a), the author has introduced an empirical mode of decomposition for signal decomposition and extracted high order statistical features for sleep staging analysis.

Sharma et al. (2019a) used wavelet function and IIR filters for signal segmentation and artifact removals. To analysis, the sleep patterns of the subjects, three different categories of features were extracted such as time and frequency domain features and non-linear features.

Sharma et al. (2017) introduced the iterative filtering techniques for EEG signal segmentation. The obtained amplitude and frequency modulated components from iterative filtering decomposition used to extract the higher-order statistical features for discriminating the sleep characteristics.

Şen et al. (2014) proposed discrete wavelet transform techniques for signal enhancement and extracted four different categories of features such as time, nonlinear, frequency, and entropy-based features. The entire study was executed from 5160 epochs of 30 s length.

Memar and Faradji (2018) extracted the time-domain, non-linear, and entropy features from eight EEG signal sub-bands such as alpha, delta, theta, gamma, beta1, beta2, gamma1, and gamma2. The total study was conducted with 142,391 epochs of three different sleep datasets.

Tian et al. (2017) extracted multi-scale entropy properties from EEG signal for characterizing the signal in multiple temporal scales manner. The study was implemented with total epochs of 18,248 of 30 s length from 10 sleep disordered and 10 healthy subjects.

Alickovic and Subasi (2018) used multi-scale principal component analysis and the informative features are extracted from signal sub-bands using discrete wavelet transform techniques. Twenty subjects participated in this ensembling sleep staging analysis.


Different classification models obtained during sleep staging

Obayya and Abou-Chadi (2014) considered single-channel EEG signal as input to identifying sleep disorders and selected subjects for this experiment work limited between 35 and 50. Here authors have obtained wavelet concept techniques for feature extraction and classified the selected features using the fuzzy algorithm. The classification model provided 85% accuracy.

Zhu et al. (2014) introduced sleep analysis through horizontal and difference visibility graphs from input signals and finally, extracted properties from the input signals are forwarded to SVM classifiers for classifying multiple stages of sleep stages and their final accuracy was 87.50% for two-state sleep stage classification problems.

Diykh et al. (2016) introduced the concept of structural graph similarity and the experimental work completely based on EEG signals. The model obtained a classification accuracy of 95.93% using the SVM classifier.

Sriraam et al. (2016) obtained multiple-channel of EEG signals from ten healthy subjects. The extracted features processed through a multilayer perceptron feedforward neural network and the overall accuracy with 20 hidden units were reported as 92.9% and subsequently for 40,60,80 and 100 hidden units in MLP, it was reported as 94.6,97.2,98.8, and 99.2, respectively.

Silveira et al. (2016) used Discrete Wavelet Transform (DWT) for signal segmentation. The extracted features were applied to a random forest classifier and overall accuracy was reported as 90%.

In Hassan and Subasi (2017) the author obtained the bootstrap aggregating concept for classifying the sleep stages. The proposed model obtained two public sleep datasets such as the PhysioNet Sleep-EDF dataset and Dreams Subjects dataset and the model reported accuracy of 92.43% for two-class sleep stages classification.

Gunnarsdottir et al. (2018) have designed an automated sleep stage scoring system with overnight PSG data and the extracted properties were classified through DT classifiers. The overall accuracy for test set data was reported as 80.70%.

Memar and Faradji (2018) considered 25 sleep suspected subjects and 20 healthy subjects for experimental purposes. The extracted features validated through the Kruskal–Wallis test and applied random forest classifier and achieved an overall accuracy of 95.31%.

In Braun et al. (2018) the author proposed a portable and effective sleep scoring system, in which he has experimented on the combination of features extracted from EEG signal and classifiers. He designed the system in such a manner that, the proposed research achieved the best classification accuracy by considering fewer frequency domain features, and the overall accuracy reported as 97.1% for the two-state classification problem.

In Huang et al. (2019), the author considered combinations of multi-variate signals for sleep scoring and extracted multiple features from both time and frequency domain features. The selected features are forwarded into SVM classifier and the reported accuracy with input single-channel EEG is 92.04%.

1.2 Contribution

It has been observed from the literature review that most of the researchers focused on accurate identification of different deployed sleep staging algorithms (O’Reilly and Nielsen 2014; Sharma et al. 2019b; Memar and Faradji 2018; Tian et al. 2017; Huang et al. 2019). The major influence to achieve high accuracy on sleep staging is selecting suitable feature parameters. It has been noticed that many of the papers did not specify this subject in detail, most of the sleep staging studies are based on signal characteristics parameters. It has been found that most of the sleep studies have been executed based on only one health conditioned subject so that sometimes it has been seen that the model may not be perfect for different categories of subjects. All these above-mentioned challenges we have addressed in our proposed sleep staging study.

In the present research on sleep staging, we propose an artificial intelligence-based automated sleep staging system using EEG signals from the subject with different health conditions subjects. The main important part of this work is the designed ensemble learning stacking model algorithm for improvement on the sleep staging classification accuracy for the five-to-two sleep states classification tasks. The proposed research completed majorly in five stages such as signal preprocessing, feature extraction, feature screening, classification algorithm, and finally the comparative analysis in between obtained different categories of subjects. The novelty of the proposed is the analysis of the sleep behavior of the subjects through proper feature screening techniques, which determines the statistically significant difference among the sleep states. Those features are selected further in the classification task, which is successfully screened through obtained feature selection techniques. The most important part of this research contribution is to propose a new ensembling learning classification model, called a stacking model for sleep staging analysis. The performance of the proposed methodology performed well for the multi-class sleep stages classification task incomparable to the existing sleep studies related to various performance evaluation metrics.

1.3 Structure of the paper

Section 2 presents briefly the proposed methodology including experimental data preparation, data preprocessing, feature extraction, and feature screening. In Sect. 3, we describe the classification algorithms used in this paper for sleep staging evaluation. Section 4 discusses the obtained experimental results from the proposed methodology from three subgroups of subjects' sleep recordings. In Sect. 5, we briefly discuss our proposed methodology results, advantages, and limitations and make a result analysis with the existing the published methods. Section 6 ends with concluding remarks with future work descriptions.

2 Materials and methodology

In this part, we presents the complete layout of the proposed multiple sleep staging classification tasks using ensemble learning stacking model. The main idea of this research work is to improvement on sleep staging accuracy with input of suitable screened features for multiple-class sleep states classification. The complete flowchart of this proposed methodology is shown in Fig. 1.

Fig. 1
figure 1

Proposed flowchart of multiple sleep staging classification tasks using ensemble learning stacking model

Full size image

The complete and integrated implementation process of the proposed ensemble learning stacking model (Fig. 1) is described in Algorithm 1. In this algorithm, we briefly presents the individual procedure for each experimental phases such as (1) data prepartion, (2) data pre-processing, (3) feature extraction, (4) feature selection, and finally (5) proposed ensemble learning stacking model.

figure a
figure b

2.1 Experimental data

In this study, the authors use three categories of subjects with different medical conditions. The recorded data were collected from the ISRUS-Sleep dataset, which was specifically for sleep studies. The data collection has been done at the Sleep Medicine Centre of the Hospital of Coimbra University (CHUC) (Khalighi et al. 2016). The first subsection includes 100 subjects, the second subsection includes 8 subjects, and finally, the third subsection includes information from 10 healthy subjects' sleep recordings. The data acquisition process is conducted in an 8–9 h full night PSG test at healthcare facilities. Signals are sampled at 200 Hz, and each epoch length is considered a timeframe of 30 s according to the AASM standard.

In this study, the authors used the C3–A2 channel for computing sleep stage classification. We considered 4 subjects one session sleep recordings from subgroup-I (SG1), 4 subject’s two-session recordings from subgroup-II (SG2), and 4 subjects from subgroup-III (SG3) with one session recordings. The SG1 (Subgroup-I) consists of 3000 epochs, while SG2 (Subgroup-II) contained 6000 epochs, and finally SG3 (Subgroup-III) consists of 3000 epoch and each epochs size is 30 s length. The entire sleep staging process was executed with all these epochs according to AASM standards. The numbers of epochs distributed among into the different sleep stages are presented in Table 1.

Table 1 Number of 30 s epochs from each sleep stage of ISRUC-sleep dataset records
Full size table

2.2 Signal preprocessing

This process starts with the segmentation of the raw signal from the three different medical conditioned subjects from the ISRUC-Sleep dataset. Each epoch of length is the 30 s and each 30 s epochs contains 6000 sample points, as the sampling frequency is 200 Hz. As we have already discussed that sleep recordings are highly complicated and random, sometimes it has found that some of the irrelevant signal information is overlapped with the brain EEG signals. Majorly, the artifacts like eye blinks and muscle movements signals are composition within the originally recorded signals. These contaminated signals give inaccurate analysis toward sleep staging classifications. To remove these artifacts from the original sleep recorded signals, we have proposed wiener filter techniques (Borowicz 2017). It is one of the parametric techniques, which supports removing the unwanted forms of noises from the original brain EEG signal. This filter mechanism is completely based on statical approaches, and it minimizes the mean square error between the estimated recording signals and the desired signals.

2.3 Feature extraction

It is one of the important step during the automated sleep staging analysis based on the EEG signals because the recorded signals are the compositions of various signal sub-bands. To analysis the changes sleep behavior from each sub-bands, extracting the features has been the best way toward recognizing the changes sleep characteristics over the individual sleep stages. The extracted signal properties may easier during discrimination the sleep patterns and alternatively it also put impacts upon the sleep staging classification performances. In the recent research progress, most of the contributions majorly focused on the three different categories of feature for signal analysis such as time domain, frequency domain, and nonlinear features (Yasoda et al. 2020; Ahmad et al. 2018; Gupta et al. 2015; Al Ghayab et al. 2019; Al-Janabi et al. 2019; Souri et al. 2020; Schetinin and Schult 2006). In this work, we obtain both time and frequency domain features. In this work, we obtain both time and frequency domain features. The time-domain analysis helps to retrieve information of fluctuating sequences of the signal and detects the epileptiform discharges. It’s very important for extracting the frequency domain features for sleep scoring to discriminate the sleep characteristics related to the individual sleep stages from EEG signal because it provides the changes of delta (δ) rhythm, theta (θ) alpha (α) rhythm, sigma (σ) rhythm, and beta (β) rhythm patterns in the EEG signals in a different frequency ranges. In this proposed research work, we extracted 28 features which include (1) 13 time-domain based features, (2) 15 frequency domain based features. The extracted different features in this proposed study and their descriptions are summarized in Table 2, was performed an epochs of 30-s length of the filtered EEG signals.

Table 2 Short explanation of the extracted features for this proposed study
Full size table

After extraction of features from respective input channels, we have applied normalization techniques before being obtained to feature selection techniques, and here we use to zero mean and unit variance, and a normalized feature vector is generated. Generally, it supports to increase the performance of the system.

2.4 Feature screening algorithms

Next to feature extraction, it is also important to select the suitable features because it has found that sometimes all the extracted features may not be relevant for all the subject cases and alternatively it directly also affects the performance of the classification models. The main intention behind considering the feature screening is to find out the most commendable features which help to discriminate the changes in sleep characteristics in the different sleep stages. Here we obtained the ReliefF feature selection algorithm for identifying the relevant features. It is one of the individual feature filtering selection methods, in which the quality and relevance of the features to be measured (Robnik-Šikonja and Kononenko 2003). The essential concept behind this algorithm is to select highly commendable features which help to discriminate the sleep behavior of the subject. As an output, this algorithm assigned a weight to individual input features according to their relevance. It determines how far the features are most discriminate to different instances amounts to different sleep stages. It generates a weight for each feature which is in the ranges from − 1(worst) to + 1(best) score. The larger the weight of the feature, the higher the association between the features and sleep stages. The main advantage of this algorithm is well managed with noisy and unknown data.

2.5 Proposed ensemble stacking model

Ensemble techniques are one of the ML approaches for improving the model by combining several models (He et al. 2019). The main advantage of this work is to reduce the variance and bias factor with the proposed stacking algorithm. It helps to increase the accuracy of the model and reduce the variability of the prediction. The proposed ensemble learning stacking model, which is in form of two-layer architecture. The first layer data and model to be treated as base-layer data and models, respectively, similarly, the second layer cross-validated data and models to be considered as meta-layer data and model, respectively. It adopts the parallel structure since the base-layer models are trained independently from each other. The base-layer predictions become input for the meta-layer, and finally, the meta classifier delivers the final predictions.

The most important advantage of the proposed model can analyze the complex sleep patterns in the meta-layer incomparable to the other parallel ensemble learning approaches. Additionally meta-layer also helps to recognize the consistently mis-predicts samples from the feature space of the base-layer classifiers due to incorrect learning and select the suitable base-layer models which are more suitable for that specific feature space. The same problems may not possible using bagging and boosting techniques. In this proposed stacking algorithm, we also use the cross-validation techniques for preventing the information leakage from the base-layer models to the meta-layer model.

3 Experimental results

The entire experiment of this research work is coded and implemented carried using the MATLAB R2017a software for signal preprocessing, feature extraction, and feature screening and also we obtained the Scikit-machine learning tool for XGBoost and mlxtend library for stacking algorithms [97] in Python running on a personal laptop with an Intel Core™ i3-4005U CPU 1.70 GHz, 2 core(s), 4 logical processors, 4 GB RAM and Windows 10 operating system. To evaluate the performance of the proposed method, a set of experiments were conducted using subjects with three different categories of medical conditions. The first experiment was conducted with the subject who was affected with different types of sleep problems and their one-session recordings considered the experiment, the second experiment carried with subjects having sleep problems and includes two different session recordings, which were recorded on two different dates. Finally, the third experiment was conducted with subjects who were completely healthy and no prior types of diseases related to sleep. In this research work, we have only considered single-channel of EEG signal as C3–A2 for extracting the sleep behavior of the subjects, according to recent sleep staging, it has found that the C3–A2 channel is most effective in analyzing the brain behavior for sleep studies in terms of classification accuracy performances because it provides central part of the brain information with related to analysis the brain-behavior during sleep (O’Reilly and Nielsen 2014; Hassan and Bhuiyan 2016a, b; Lei et al. 2018; Sharma et al. 2017, 2019a, b; Zhu et al. 2014; Ronzhina et al. 2012; Spriggs 2014; Şen et al. 2014; Memar and Faradji 2018; Tian et al. 2017; Alickovic and Subasi 2018). Hence, we have used the C3–A2 channel of EEG signal recordings for sleep staging for all the three categories of subject data in our proposed work.

In this research work, we have conducted four individual experiments for sleep staging. The first three experiments were conducted with three different subgroups subjects’ sleep recordings with help of base layers classification algorithms such as SVM, DT, KNN, and RF. The final experiment was our proposed ensemble learning stacking model, where we used another layer for learning the model, called as meta classification layer. We used the XGBoosting algorithm in the Meta classifier layer, which considered input as base layers predictions and made final decisions on sleep scoring. In the proposed sleep study, we conducted multi-class sleep stages classification: CT-2 (Wake vs. Sleep), CT-3 (Wake vs. NREM vs. REM), CT-4 (Wake vs. N1 + N2 vs. N3 vs. REM) and CT-5 (Wake vs. N1 vs. N2 vs. N3 vs. REM).

3.1 System performance evaluation metrics

The proposed study provides an in-depth analysis based on a comparative analysis of multiple different subjects with different session recordings obtained for sleep stage scoring analysis. For that reason, the authors have used multiple evaluation metrics to analyze the performance of the proposed sleep stage classification method.

This study considers four criteria such as the classification accuracy (Raza et al. 2016), recall (Bajaj and Pachori 2013), specificity (Hsu et al. 2013), precision (Zibrandtsen et al. 2016), F1Score (Berry et al. 2014), and kappa score (Statistic and in Reliability Studies: Use, Interpretation, and Sample Size Requirements 2005).The confusion matrix is used to evaluate the results obtained from classification algorithms. The authors have analyzed the information about the actual and predict score achieved by the algorithms used.

3.2 Performance of sleep staging using Subgroup-I (SG-I) dataset

In this experiment, we considered with single-channel C3–A2 of EEG signal recordings of four sleep-disordered subjects. All four subjects have diagnosis with sleep syndrome disease along some of them also suffered from insomnia and other types of sleep problems. The selected features are forwarded into the obtained four base layers classifiers such as SVM, DT, KNN, and RF for initial predictions on sleep staging. The classification accuracies (CAs) achieved for the two-five sleep states from base classifiers are observed. Table 3 presents the summary of the reported classification accuracy results (CAs in %) from obtained classifiers.

Table 3 Summary of the accuracy results for different classifiers with SG-I dataset
Full size table

It has observed from Table 3 that, the highest classification accuracy reported for CT-2 (99%), CT-3 (97.97%), CT-4 (97.77%) and CT-5 (96.33%) through RF classifier, similarly the lowest accuracy for CT-2 (93.87%), CT-3 (89.27%), CT-4 (85.07%) and CT-5 (84%) noticed from DT classifier. The highest accuracy reported RF classifier confusion matrix results for CT-5 is presented in Table 4.

Table 4 Confusion matrix for CT-5 using Random forest classifier using SG-I dataset
Full size table

3.3 Performance of sleep staging using Subgroup-II (SG-II) dataset

In this section, we have considered four subjects and their two different session sleep recordings, which were recorded on two different dates from the enrolled subjects. The subjects were under-diagnosis of roncopatia disease and also some of them were affected with various types of sleep problems. The same channel was used for signal acquisition and extracted the same features from the processed EEG signal. The same obtained classification algorithms used here and the reported classification accuracy performance from all the classifiers are shown in Table 5.

Table 5 Summary of the accuracy results for different classifiers with SG-II dataset
Full size table

From classification performance results, it has been found that the RF classification model performs well for two-class to five-class classification tasks. The reported accuracy for CT-2 (98.62%), CT-3 (90.2%), CT-4 (89.71%) and CT-5 (89.1%). The KNN model achieved very low accuracy in terms to sleep scoring for CT-2 (90.33%), CT-3 (88.77%), CT-4 (84.54%), and CT-5 (82.15%). The confusion matrix for CT-5 using RF classifier is presented in Table 6.

Table 6 Confusion matrix for CT-5 using Random forest classifier using SG-II dataset
Full size table

3.4 Performance of sleep staging using Subgroup-III (SG-III) dataset

In this experiment, we have obtained completely four healthy controlled subjects' sleep recordings, who have not affected any types of sleep problems during sleep hours. The performance achieved from base learning classifiers using the SG-III dataset is presented in Table 7.

Table 7 Summary of the accuracy results for different classifiers with SG-III dataset
Full size table

It has observed from Table 7 that, the highest classification accuracy performance reached with random forest classification model for CT-2 (99.10%), CT-3 (98.07%), CT-4 (97.93%), and CT-5 (98.13%). The confusion matrix for reporting higher accuracy with RF model is presented in Table 8 with respect to CT-5.

Table 8 Confusion matrix for CT-5 using Random forest classifier using SG-III dataset
Full size table

3.5 Performance of sleep staging using proposed Ensemble stacking algorithm using subgroup-I/II/III datasets

Finally, we have deployed our proposed ensemble techniques for sleep scoring, using the integration of the base layers classification model. In this study, we have used four classification techniques such as SVM, DT, KNN, and RF for taking the first predictions on sleep scoring. This computed first layer prediction output is considered as to the second layer, called as meta classification layer. We obtained an XGBoosting classification algorithm in the second layer, which integrates the previous layer's predictions and generates the final decisions on sleep staging. The confusion matrix of this proposed ensemble learning stacking model using SG-I, SG-II, and SG-III datasets for five sleep states classification problems are presented in Tables 9, 10, and 11, respectively. The final classification performance of this proposed model with input of SG-I, SG-II, and SG-III datasets are presented in Table 12.

Table 9 Confusion matrix for CT-5 using proposed ensemble learning stacking model using SG-I dataset
Full size table
Table 10 Confusion matrix for CT-5 using proposed ensemble learning stacking model using SG-II dataset
Full size table
Table 11 Confusion matrix for CT-5 using proposed ensemble learning stacking model using SG-III dataset
Full size table
Table 12 Performance of five-state sleep (CT-5) staging based on ensemble learning stacking model with SG-I, SG-II, and SG-III datasets
Full size table

It has found from Table 12 that, the proposed ensemble learning stacking model performed excellent with related to five sleep states classification problem in comparable to base layers classification algorithms such as SVM,KNN,DT and RF. The proposed stacking model reported classification accuracy of 99.34%, 96.30%, and 98.50% for SG-I, SG-II, and SG-III datasets, respectively.

3.6 Summary of results

In this paper, we were used three subgroups data of ISRUC-Sleep for sleep staging. In all the experimental part conducted for this study using single-channel EEG signal. The first three experiments conducted based on base layers learning classification model and the final experiment executed with the proposed stacking model. Table 13 presents the summary results that was obtained through different classification models of base layers and ensemble learning of stacking model with the three different health categories of subjects sleep recordings using SG-I, SG-II, and SG-III datasets.

Table 13 Summary results of classification accuracy for various classification models and subgroups datasets
Full size table

The highest sleep staging was obtained using ensemble learning stacking model from all the three subgroups dataset. The highest classification accuracy was reported as 99.34% for five sleep states classification problem with SG-I dataset.

4 Discussion

A machine learning (ML) based automated sleep scoring is suggested for the classification of multiple sleep stages. Several experiments were executed on three subgroups datasets to validate the potency of the proposed methodology. The proposed sleep staging methodology can automatically learn high level from the single-channel EEG signal directly.

4.1 Hypothesis and limitations of the proposed model

The proposed automated sleep staging results indicate that the ensemble learning stacking model reported improvement in classification accuracy compared to the other earlier automated sleep staging and other ML methods. In comparison with the other contributed sleep staging classification problems, we mainly focus on four things that help to improve the classification accuracy of sleep staging: 1.preprocessing the signal using wiener filter techniques, 2.feature screening algorithm using ReliefF weight techniques, and 3.ensemble learning classification techniques.

One of the major advantages of this proposed model is using three different medical conditioned subjects with their different sleep recordings. One of the other major focused things in this work was the proper screening of the features by considering the ReliefF weight algorithm and eliminating redundant features using Pearson correlation techniques. In this research work, we have used wiener filtering techniques which support a lot toward reducing the artifacts from the input channel. Our proposed methodology was executed with 12,000 epochs, each epoch of 30 s length for two-five sleep stages classification with three subgroups datasets.

The major difficulty with PSG signal is made the more uncomfortable situation for subjects during sleep recordings due to more connecting electrodes and rigid placement postures (Al-Janabi et al. 2019). Sometimes this uncomfortable scenario and possible body movements during sleep recordings degrade the quality of input signals. Finally, we have obtained the ensemble learning model which combinations of two-layers of classification models, first layer is considered as the base learning layer, and the second layer is called a meta-learning layer. The final decisions to be obtained from the meta-layer classification model. Even though our proposed framework achieved very good accuracy incomparable to other state-of-the-art works. Till, there is more way for us to improve our model to reach classification accuracy closer to 100% and this result we can achieve by overcoming the class imbalance problem. Also, we will use deep learning techniques, in which automated feature learning helps us to overcome variations on feature influence problems and supports to reach the high classification accuracy on sleep staging. The class imbalance problem solves by implementing data augmentation concepts using deep learning techniques.

4.2 Comparisons of the proposed research work performances with other single-channel sleep staging classification

The authors have compared the proposed system performance with other studies available in state of the art. Therefore, the authors have selected studies with similar datasets according to our proposed study and based on a single channel. In Table 14, the features used in the proposed research work are compared to others used by related works using single-channel EEG signals of ISRUC-Sleep dataset. The comparisons with other similar research proposals available in the literature must take into consideration the use of single-channel EEG, different features and classification models is presented in Table 15.

Table 14 Comparisons of CAs (in %) of our proposed model with other state-art-of the techniques used same features and datasets
Full size table
Table 15 Comparisons of sleep staging accuracy (%) between the proposed methodology and the recent methods
Full size table

Tables 14 and 15 present a comparison of the overall accuracies presented by the proposed methods and the available studies in the literature. The results reported, but the authors using the ensemble learning stacking model for SG-III, SG-II, and SG-I present the highest results when compared with the other methods available in the literature. The proposed research work has obtained best accuracy results for five-sleep (CT-5) classification. We achieved an overall accuracy of 99.34% with SG-I, 90.8% with subgroup-II and 98.50% with SG-III subjects through stacking model. These classification results state that the proposed research work provides a significant contribution to the field of automated sleep stage classification and can support the decision-making by providing an efficient healthcare model to diagnosis the sleep-related diseases.

5 Conclusion and future directions

This paper has presents an ensembling learning stacking model under AASM sleep scoring rules for two–five sleep states classification using a single channel EEG signals. The proposed methodology has been analyzed 12,000 epochs composed of three different medical conditioned subject datasets.

There are three main important contributions in this research study. The first one has obtained wiener filter techniques for eliminating the various artifacts that exist in the input signals and followed by we have extracted numerous features from linear and nonlinear features. This set of features supports the analysis of sleep EEG parameters and their characteristics. It has been observed that multi-feature extraction improves sleep staging accuracy. Secondly, the proposed research work obtained feature screening techniques, which directly useful for identifying the most relevant features from extracted feature vectors. Additionally, we also eliminate the redundant features from selected relevant features using Pearson correlation analysis, which helps us to select the suitable features for classification tasks. Thirdly, this proposed research work establishes an ensemble learning model, which integrates multiple classification models implements in two layers. In the first base layers, we obtain SVM, DT, KNN, and RF classifiers, and the second layer contained XGBoosting techniques. The proposed stacking model reported high recognition rates for multiple sleep staging classification tasks.

The results show that this study can provide an effective mechanism for handling different health conditions of the subjects with high accuracy. The authors have also compared the proposed research with other similar studies available in the literature to show better performance and state the present contribution. Moreover, the proposed system presents high performance through all the categories of subjects with different medical conditions. Finally, this study uses the ISRUC-Sleep three subgroups dataset, which is a useful information source for a sleep evaluation.

Furthermore, the authors use shorter epochs lengths (the 20 s, 15 s, 10 s, and 5 s) to the development of a real automated sleep staging system. Consequently, the authors aim also to consider a high number of clinical sleep data. In particular, the future work will include the different sleep problems patients to evaluate the performance of the proposed method work for higher accuracy using deep learning techniques.