Tree-Based Machine Learning Techniques for Automated Human Sleep Stage Classification

In this part of the manuscript performance of the proposed manuscript compared with similar studies up to date by using number of classes (Wake (W), rapid eye movement (REM), non-rapid eye movement (NREM)), PSG device input type and dataset type. The proposed work offers many advantages such as having its own database, using multi-channel inputs, and having high accuracy. Table 1 summarizes a list of comparisons between the proposed work and recent similar works. To explain the presented works importance following works chosen which employs different inputs and different techniques to show the literature gap.

Shahin et al. [35] realized a deep learning approach by using single and two EEG inputs on their own dataset to classify sleep stages. When compared with manual assessments, an NREM+REM based classifier had an overall discrimination accuracy of 92% and 86% between two groups using both two and one EEG channels, respectively in 2017. This study also proves that the accuracy has a directly relation with channel input numbers. Li et al. [36] classifies three sleep stages (Wake-REM-NREM) from a single lead ECG using beat detection, cardiorespiratory coupling in the time frequency domain and deep convolutional network (CNN). Their obtained accuracy was 85.1% when application was verified on MIT-BIH PSG database. DNN was used for the classification of the sleep stages into W, REM and NREM sleep stages by Wei et al. [37] in 2018. They applied the sleep stage stacked autoencoder to constitute a 4-layer DNN model. To test the accuracy of their method, eighteen PSGs from the MIT-BIH Polysomnographic Database were used. They obtained an accuracy of 77% and a Cohen’s kappa coefficient of about 0.56 for the classification of W, REM, NREM. Another deep learning-based sleep stage classification work made by Chambon et al. [38] in 2018. They employed 6 EEG, 2 EOG and 3 EMG channels of publicly available MASS dataset. Their accuracy for five sleep stage classification was around 87%.

Malafeev et al. [39] developed machine learning algorithms for sleep classification: random forest (RF) classification based on features and artificial neural networks (ANNs) working both with features and raw data. They tested their methods and achieved accuracy of 90% when on 5 stage classifiers. Zhao et al. [40] employs SVM for classification of sleep stages on EDF dataset by using EEG input. Their accuracy is 85.93% when this technique was applied to EDF database in 2019. A flexible deep learning model is proposed using raw PSG signals by Yildirim et al. [41]. A one-dimensional convolutional neural network (1D-CNN) is developed using electroencephalogram (EEG) and electrooculogram (EOG) signals for the classification of sleep stages. The performance of the system is evaluated using two public databases (sleep-edf and sleep-edfx). The developed model yielded the highest accuracies of 98.06% and 94.64 for 2 and 3 classes, respectively. Ravan and Begnaud [42] used quantitative features obtained from electroencephalography in 2019 for awake-Ligth sleep+REM-deep sleep classification. They first developed a new automatic sleep-staging framework that consists of a multi-class support vector machine (SVM) classification, based on a decision tree approach. To train and evaluate the performance of the framework, they used polysomnographic data of 23 healthy subjects from the PhysioBank database where the sleep stages have been visually annotated. After that, they used the trained classifier to label the sleep stages using data from 22 patients. Their overall accuracy is 90% at the related study when PZ-Oz channel was employed. Jeon et al. [43] proposed a novel end-to-end classifier based on a multi-domain hybrid neural network (HNNmulti) approach consisting of a convolutional neural network and bidirectional long short-term memory for automatic sleep staging with pediatric scalp EEG recordings in 2019. Their HNN-based classifier yielded the best performance metrics using 30-s time series in combination with an instantaneous frequency using a 19-channel, three-stage classification with overall accuracy, F1 score, and Cohen's Kappa, equal to 92.21%, 90%, and 88%, respectively. Huang et al. [44] proposed a multi-channel signal adding method for sleep staging in 2020. Overall performance was obtained while using the proposed method, and in six-state sleep staging. The highest overall classification accuracy of 96.53% was obtained based on the superposition of two EEG signals. Chriskos et al. [45] realized a study for automatic sleep staging by employing CNN and cortical Connectivity images. When they apply Synthetic Minority Oversampling Technique (SMOTE) technique for the classification REM-NREM stages, they obtained an accuracy of 99.85%. Santaji and Desai [46] proposed an efficient technique for sleep stage classification based on Electroencephalogram (EEG) signals analysis using machine learning algorithms by considering 10 seconds of EEG epochs. EEG signals were filtered and decomposed into frequency sub-bands using band-pass filter. Statistical features are extracted and trained with Decision Tree, Support Vector Machine and Random Forest algorithms with different testing dataset percentage. Results show Random Forest algorithm achieves 97.8% of accuracy for NREM-REM classification.

Zhang et al. [47] collected data from 294 studies and developed a model by using DNN to obtain higher accuracy. Their accuracy value is 96.02% in binary classification (as N1 and N2, W and REM ETC). However, the overall classification accuracy decreases to 83% in 5 classes classification. Satapathy et al. [48] proposed a study aims to develop a new automated sleep staging system using the brain EEG signals. Based on a new automated sleep staging system based on an ensemble learning stacking model that integrates Random Forest (RF) and eXtreme Gradient Boosting (XGBoosting). Their model achieved an accuracy of 91.10% in S-EDF dataset. Grieger et al. [49] proposed a classification system based on a simple neural network architecture that scores the classical stages as well as pre-REM sleep in mice. Their accuracy for NREM, W, REM was 97%. Satapathy and Loganathan [50] proposed automated sleep staging system followed four basic stages: signal preprocessing, feature extraction and screening, classification algorithms, and performance evaluation. In that work, a new method is applied for signal preprocessing, feature screening and classification models. With that proposed model they obtained an accuracy of 99.34% on SG-I dataset for 5 stage classification. Another high accuracy work done by Satapathy and Loganathan [51]. That work develops an Automated Sleep Staging System based on Two-Layer Heterogeneous Ensemble Learning Stacking Model (ASSS-TL-HELSM) for sleep staging. Their model has maximum accuracy of 99.02% in SG-I dataset if 3 substage (W-REM-NREM) and feature selection were applied. Arslan et al. [34] proposed a 5-class model for automatic scoring of sleep stages. Classification was made using 19 sensors and as a result, 95.36% accuracy value was obtained for Extra Tree.

Abdollahpour et al. [52] proposes a new method for fusing two sources of information, electroencephalogram (EEG) and electrooculogram (EOG), to achieve promising results in the classification of sleep stages. The proposed method employs transfer learning at the training stage of the model to accelerate the training process of the CNN and to improve the performance of the model. The proposed algorithm was used to classify the sleep-EDF and sleep-EDFX benchmark datasets. The algorithm could classify the Sleep-EDF dataset with an accuracy of 93.58% and Cohen’s kappa coefficient of 0.899. The results show that the proposed method can achieve superior performance compared with state-of-the-art studies on the classification of sleep stages. Furthermore, it can provide reliable results as an alternative to conventional sleep staging.

Fraiwan and Alkhodari [53] examines the application of a long-short-term memory (LSTM) learning system for the purpose of automatic sleep stage scoring in neonates. The research employed a dataset of 5095 sleep stage signals that were obtained from electroencephalogram (EEG) recordings conducted at the University of Pittsburgh. The Pediatric Neurology Department of Case Western Reserve University enlisted the expertise of a medical doctor to annotate the sleep stages of neonates. Specifically, the doctor identified three distinct sleep stages-awake (W), active sleep (AS), and quiet sleep (QS)-in 60-second epochs. The signals underwent pre-processing procedures, including normalization and filtering. The signals obtained were partitioned into four-fold, six-fold, and 10-fold cross-validation schemes. The bidirectional LSTM network classifier, which has been constructed with predetermined training parameters, is used to execute the training and classification procedures. The algorithm that was formulated underwent an evaluation process, which was accompanied by a comprehensive summary table that presents the findings of this investigation as well as those of other contemporary research endeavors. The present investigation attained notable levels of Cohen's kappa (κ), accuracy, and F1 score, specifically 91.37%, 96.81%, and 94.43%, respectively. The confusion matrix indicates that the true-positive percentage achieved an overall value of 95.21%. The algorithm that was developed demonstrated favorable outcomes in the context of automated scoring of neonatal sleep signals for sleep stage identification. Subsequent research endeavors will entail enhancements to the classifier’s overall accuracy through the utilization of LSTM architecture and improvements to the training parameters .

Eldele et al. [54] introduces a new deep learning framework named AttnSleep, which employs attention mechanisms for the classification of sleep stages based on EEG signals obtained from a single channel. The architectural design comprises three primary components, namely feature extraction, temporal context encoder, and classification. The module for feature extraction relies on a convolutional neural network (CNN) that operates at multiple resolutions, known as the multi-resolution CNN (MRCNN). Additionally, the module employs adaptive feature recalibration (AFR) to enhance the quality of the extracted features. This approach allows for the extraction of both low-and high-frequency features and models the interdependencies between them, resulting in improved feature quality. The Temporal Context Encoder (TCE) utilizes a multi-head attention mechanism to effectively capture temporal dependencies within the extracted features. The module responsible for categorizing sleep stages employs a fully connected layer for classification purposes. The AttnSleep model, as proposed, exhibits superior performance compared to contemporary techniques, as evidenced by various evaluation metrics.

The paper proposes an efficient technique for sleep stage classification based on Electroencephalogram (EEG) signals analysis using machine learning algorithms by considering 10 s of epochs. The EEG signals are filtered and decomposed into frequency sub-bands using a band-pass filter. Statistical features are extracted and trained with Decision Tree, Support Vector Machine and Random Forest algorithms with different testing dataset percentage. Results show that the Random Forest algorithm achieves 97.8% of accuracy. The paper also mentions that PSD and ERP are well-established methods for analyzing EEG signals to classify sleep stages. In this study, PSD and ERP plots are derived using EEGLAB, which is a graphical user interface that permits users to intuitively process the data for better sleep stage classification [55].

Liu et al. [56] examines the use of polysomnography as a preeminent method for detecting sleep stages, and underscores inconsistencies in the application of these criteria by technicians. The present research suggests the implementation of an artificial intelligence (AI) system to effectively assess the dependability and uniformity of sleep scoring, and consequently, the quality of sleep centers. This was achieved through the use of an interpretable machine learning algorithm to evaluate the interrater reliability (IRR) of sleep stage annotation among sleep centers. A study was conducted on 679 patients without sleep apnea from six sleep centers in Taiwan to perform intra center and inter center assessments. Centers that may have quality issues were identified using the estimated internal rate of return (IRR). Intra center assessment revealed that the median accuracy varied between 80.3% and 83.3%, apart from a single hospital, which demonstrated an accuracy of 72.3%. During the inter-center assessment, the median accuracy varied between 75.7% and 83.3% when a single hospital was omitted from both the testing and training phases.

Haghayegh et al. [57] presents a novel deep learning algorithm that integrates Proportional Integrating Measure (PIM) and zero-crossing mode (ZCM) data to estimate sleep parameters through wrist actigraphy. The research entailed the acquisition of ZCM, PIM, and electroencephalographic (EEG) data from a sample of 40 individuals who were in good health. The algorithm under consideration demonstrated a noteworthy improvement in specificity compared to the existing algorithm while exhibiting a slight reduction in sensitivity for individuals suffering from sleep disorders. The inconspicuous evaluation of circadian rhythms is especially pertinent for individuals with neuropsychiatric disorders linked to sleep disruptions, such as major depressive disorder or cognitive decline. The manuscript additionally examines approaches to mitigate the issue of incomplete data through the optimization of DHT deployment and the incorporation of patient viewpoints in the research framework. Furthermore, this manuscript presents a methodological guide for establishing studies on daily life, with a specific emphasis on evaluating salivary cortisol levels. A polysomnography (PSG) study was conducted on a sample of 11 male and 9 female individuals in order to assess potential neuropsychiatric sleep disorders. Concurrently, wrist actigraphy was documented, whereby 37 characteristics were calculated for every 1-minute interval. The study involved a comparison of the prediction of PSG-derived sleep-wake states for each feature using our newly developed algorithm and four state-of-the-art algorithms. The performance of the algorithms was assessed through the use of leave-one-subject-out cross-validation. The recently developed algorithm demonstrated an accuracy of 84.9% in identifying sleep epochs and 74.2% in identifying wake epochs, resulting in an overall sleep-wake scoring accuracy of 79.0%.

Peker offers a concise survey of the literature pertaining to the categorization of sleep phases. The text delves into a comparative analysis of the proposed and current methodologies while referencing the novel contributions of the proposed approach. The present study introduces a novel approach to automatic sleep scoring utilizing single-channel electroencephalography (EEG) signals. This approach is a hybrid machine-learning method that incorporates complex-valued nonlinear features (CVNF) and a complex-valued neural network (CVANN). The nomenclature assigned to the proposed technique is CVNF CVANN. The method under consideration demonstrated accuracy rates of 91.57% and 93.84% in accordance with the R&K and AASM standards, respectively. These results suggest that the method has the potential to be effective in the context of sleep scoring. The present study introduces a facial video database and its corresponding acquisition process, comprising 31,500 video clips featuring 100 distinct individuals hailing from 20 different countries [58].

As briefly given here and other review studies [59, 60] our works differentiated from related works by having own dataset and applying all channels as input. Since sleep experts uses all PSG channels during sleep stage scoring, we wanted to all PSG channels of the device as input which leads to 5 million data for every patient which is occupied from 800 epochs of 30 seconds approximately. For increasing reliability, sleep data had been used which means 250 million records. Machine learning algorithms applied and obtained results presented in detail. The maximum accuracy obtained as 98.8% when machine learning algorithms applied in this work. Although obtained accuracy is lower than the ones that uses publicly available dataset with single or dual channel data, we claim that the proposed work will give an idea to the researcher who want to make similar studies by using own dataset with all PSG inputs. In order to explain the literature gap, Table 1 is prepared by giving classes, PSG inputs, dataset and size, engineering techniques and accuracy tabs as below. The proposed work differs from related works by using own dataset and applying discrete signal pre-processing and processing on it to achieve recognize the sleep stage with nearly 100% accuracy. When we look at the current state of the art, similar works employs EEG, ECG, or whole PSG inputs to differentiate the sleep stages generally by using DNN, machine learning or CNN methods on publicly available datasets. This work is motivated by that literature gap. Another point of literature gap which can be seen in the Table 1, the number of individuals in approximately half of the databases used in similar studies is less or the same amount than the number of individuals in ours. In studies [50, 51], were conducted on 3 groups and the number of individuals in two of them was less than 50. The number of individuals in the two references is around 60, not much different from 50. As a result, it is seen that the almost 70% of similar works have similar number individuals as in this work individuals in the databases used in 70% of similar studies is similar to that in our database.

A summary table of the studies on the subject is presented in Appendix A.

After the data was preprocessed, classifier and features selected, it was tested using machine learning techniques. The results are given in detail in this section and has compared each other.

4.1 Dataset

The data set was provided by Yozgat Bozok University, Department of Chest Diseases, Sleep Laboratory. The dataset consists of PSG signals collected from 50 subjects and recorded with the Philips Alice PSG device. PSG signals are collected over a night that lasts approximately 8 hours, via sensors attached to the subjects’ bodies and operating at various frequencies. In this study, the signals collected from 19 different channels are reduced to 16-channel signals by using feature selection and the dataset is created. These data are grouped into epochs, each representing a 30-second time frame, according to AASM standards. Each epoch was evaluated by well-trained sleep specialists according to AASM standards and sleep stage scored. After the recordings classified as N1, N2 and N3 by the sleep specialist are combined under the name of NREM, sleep stage scoring is performed according to three classes: NREM, REM and WAKE.

The sensors that make up the 16 different channels do not all operate at the same frequency. Therefore, data collected at different frequencies is appropriately organized during data preprocessing, as explained in Section 3.2. According to this arrangement, an epoch consists of 6000 records, each containing data from 16 different channels. There are approximately 800 epochs in the dataset created for a subject. This means that there is 800×6000×16 values in a subject’s dataset. Although these values vary from subject to subject, they are approximately 75 million. The number of records (rows) in the dataset of 50 different subjects according to sleep stage classes are shown in Appendix B. In this study, sleep stage classification was performed on a row basis, not on an epoch basis. So, each of the approximately 800×6000=4.8 million rows are classified using approximately 75 million data in the calculations.

Number of records (rows) in the subjects’ dataset according to different sleep stages is given in Appendix B.

4.2 Performance metrics

In this study, automatic scoring of sleep stages is done per recording. Thus, it is aimed to evaluate the sleep quality and sleep disorders of the subjects. The evaluation metrics of per-record scoring, accuracy (acc), recall, precision (prec), and f1-score, are defined as follows:

Accuracy $=\frac{(T P+T N)}{(T P+T N+F P+F N)}$     (1)

Recall $=\frac{T P}{(T P+F N)}$     (2)

Precision $=\frac{T P}{(T P+F P)}$     (3)

$F 1-$ score $=2 \times \frac{(\text { Recall } \times \text { Precision })}{(\text { Recall }+ \text { Precision })}$      (4)

4.3 Experimental environment

All experiments were performed on a computing server equipped with 2 x Intel(R) Xeon(R) Silver 4114 CPU @ 2.20 GHz and 192.0 GByte random access memory. The server uses a system running Windows 10 Pro for workstations. The results were obtained using the Python v3.9.5 programming language and its libraries sklearn, pandas, numpy, seaborn, matplotlib and imblearn. The environment is particularly important in calculations regarding training and test times.

4.4 Testing results with all classifiers

In this study, it was aimed to achieve higher performance in scoring sleep apnea. For this purpose, the results obtained with 11 classifiers were as given in Table 5. Accordingly, it has been determined that tree-based models, RF, DT and ET, are suitable for this classification.

Table 5. Classification results with different algorithms

The results for sleep stage estimations using the Extra Tree, Random Forest, and Decision Tree classifiers for each of the 50 subjects are shown in Figure 5. According to these results, it is observed that the Extra Tree and Random Forest classifiers have the same performance, while the Decision Tree classifier has slightly less performance. Accuracy values are 0.950, 0.954, and 0.92 for the Extra Tree, Random Forest, and Decision Tree classifiers, respectively. Precision, Recall and F-Score values also show parallelism with Accuracy values. These results are the calculated mean values for 50 subjects. Detailed results for each classifier for all subjects are given in Appendix D. When the subjects are analyzed separately, it is seen that the Accuracy value is between 0.925 and 0.988 for Extra Tree, between 0.930 and 0.988 for Random Forest, and between 0.881 and 0.978 for Decision Tree. These results prove that the proposed model using 16 different channel data is significantly successful in solving the classification problem. Figure 6 shows the confusion matrix for the dataset where the highest classification performance is obtained with an accuracy of 0.988. As seen in the matrix, it is understood that approximately the same success can be achieved for all classes. It is also seen from the matrix that remarkably high True Positive values are obtained for all classes. This indicates that NREM, REM and WAKE sleep stages can be successfully scored with the proposed model.

In cases where the data between classes is balanced, the performance of the model obtained can be shown with the ROC curve. However, ROC curves produce optimistic results when class distributions deviate. Precision-recall curves are an alternative to ROC curves because there are differences in class distributions. In the dataset used in this study, as can be seen in Appendix B, sample imbalance between sleep stages is high. For this reason, in Figures 6-8, class-based precision-recall curves taken separately for each of the 3 classifiers are given. As can be seen in all of the curves, the area value is very close to 1.0. This suggests that the proposed provides both high precision and high recall. It proves that a complete representation of an ideal classifier is used in this work.

5.png

Figure 5. Accuracy, precision, recall and f1-measure of obtained results on extra tree (a) random forest (b) decision tree (c) classifiers

6.png

Figure 6. Precision-recall curve of DT

7.png

Figure 7. Precision-recall curve of ET

Table 6. Stratified 10-fold cross validation of tree-based models

The performance of the proposed model in this study is expressed with different metrics and measurements. However, cross-validation is needed to show that the results obtained show the same success in each random selection. Thus, it is possible to make a less biased evaluation for the proposed model. In order to make this objective comparison, Balanced accuracy and f1 weighted values were measured for 3 different tree-bases classifiers by choosing k=10 value and presented in Table 6. As can be seen when the tests are repeated by performing 10-Fold, no notable change was observed in the results. Average classification performance remained constant for both balanced accuracy and f1 weighted value. This situation reveals that the proposed model is not affected by training and test data changes and is a robust model in producing similar results.

The final evaluation expected for the results obtained is on the FP and FN numbers of tree-based models on a class basis. Accordingly, in Figures 9-11, “confusion matrices” are given for 3 classifiers. Accordingly, the proposed model can detect the NREM class with higher performance than other classes. Proportionally the lowest performance is in the REM class. It is considered that this is due to the combination of N1, N2 and N3 classes as NREM classes. This situation revealed that FP values increased with the REM class containing different data. It was observed that all 3 tree-based models gave close values to each other.

8.png

Figure 8. Precision-recall curve of RF

9.png

Figure 9. Confusion matrix of DT

10.png

Figure 10. Confusion matrix of ET

11.png

Figure 11. Confusion matrix of RF

As a result, the feature selection proposed in this study, using hyper-parameter tuning, used tree-based classifier models to solve the 3-class sleep apnea problem with high performance. The resulting model performs well independently of class and patient and produces consistent results for this problem with an unbalanced sample number based on class.

4.5 Discussion

Three sleep stage classes occupied from NREM, REM, and WAKE has been researched and machine learning model proposed in this study. High classification success was achieved by using 16 different input channels. This performance was demonstrated on a dataset of 50 different subjects. While the highest performance was achieved with 98.8% accuracy on the basis of the subject, the overall system was found to be quite successful with an average accuracy of 95.4%. Table 7 summarizes similar studies on this subject in recent years. As can be seen in this summary table, some of the studies on this subject are different from ours in that they only score 2-class, some use only EEG or both EEG and EOG channels, and some use public databases. Among similar studies scoring on a three-class basis, only one [48], using data collected from mice, outperformed ours by 99.34%. When it comes to sleep disorders in humans, the proposed model may inspire researchers.

Although the classification success of the proposed model in this study is quite high, it also has some disadvantages. The large amount of data in the used dataset increases the classification time and requires high capacity computing devices, resulting in high processing time requirements. With the solution of this problem, the proposed model can reach widespread use. It could be useful to discuss advantages and disadvantages of the paper by briefly here. First of all, the study was enhanced by a thorough data collection procedure involving a sample size of 50 participants. The dataset has been gathered within a regulated laboratory setting using established methodologies, thereby augmenting the reliability of the data. The utilization of 16 distinct input channels facilitated a comprehensive examination of the various sleep stages through the implementation of multichannel input. The incorporation of this particular approach introduced an additional level of complexity to the investigation, which plausibly enhanced the precision of the findings, demonstrating a notable level of precision with an average accuracy of 95.4% and a maximum accuracy of 98.8%. The degree of precision demonstrated is not only noteworthy in its own right but also exhibits favorable parallels to analogous investigations, indicating that the framework exhibits considerable potential for pragmatic implementation. The research provides comprehensive and tailored perspectives for each unique participant. The degree of granularity involved in the analysis aids in comprehending the distinctions among subjects and enhances the adaptability and profundity of the analysis. The efficacy of the Random Forest classifier in producing precise outcomes highlights the adaptability and resilience of the model which can effectively process diverse PSG datasets. There are also disadvantages which can be discussed together with advantages. One notable limitation of this study is its high degree of data intensiveness, as a considerable volume of data is necessary for conducting the analysis. The potential outcome of this phenomenon is an increase in computational expenses and prolonged processing durations, thereby constraining the feasibility of the model in environments with limited computational capabilities. The use of 16 distinct channels in the study may potentially augment the intricacy of the model. Potential difficulties in interpretability and implementation of the model may arise, particularly in less controlled environments or with less sophisticated devices. The study’s findings rely on the sleep stages that were evaluated by professionals specializing in sleep scoring. The presence of any discrepancies or bias in the preliminary assessments conducted by humans may have an impact on the precision of the machine learning algorithm. The generalizability of the study is uncertain because the model’s performance on datasets with fewer channels or less controlled environments is unknown despite achieving high accuracy with its specific dataset.

Table 7. Similar studies on the sleep staging in recent years

Bakır and Ulutaş developed the search strategy and conducted the literature research. Ulutaş and Köksal performed data extraction and pre-processing. Çiftçi performed manual scoring of sleep stages. Arslan et al. interpreted the evidence from a methodological and clinical point of view. Arslan supervised the execution of the study. All authors contributed to the drafting of the article and read, critically reviewed, and approved the final article.