2022 25th International Conference on Computer and Information Technology (ICCIT) 17-19 December, Cox’s Bazar, Bangladesh
arXiv:2606.07351v1 [cs.LG] 5 Jun 2026
SleepExplain: Explainable Non-Rapid Eye Movement and Rapid Eye Movement Sleep Stage Classification from EEG Signal Rafsan Jany
Md. Hamjajul Ashmafee
Iqram Hussain
Islamic University of Technology [email protected]
Islamic University of Technology [email protected]
Seoul National University [email protected]
Md Azam Hossain Islamic University of Technology [email protected]
standard for sleep stage classification (AASM). According to this standard, PSG signal recordings are classified as NonRapid Eye Movement (NREM) sleep, Rapid Eye Movement (REM) sleep, and waking (W). The American Academy of Sleep Medicine (AASM) has more recent guidelines for this (AASM) [8], [9]. The AASM rules also specify the distinctive waves for each of the five sleep phases [12]. During NREM and REM stages, the human body has to face many functional changes in both the nervous system and the body system [18]. Hormonal changes also occur in these two stages [18]. When NREM sleep goes deep, sympathetic nerve functionality decreases. There is a break in sympathetic nerve activity at some point of NREM sleep due to the short increase in blood pressure and heart rate that follows K-complexes [18]. During REM sleep, respiratory flow and ventilation change and become faster and more erratic [19], [20]. Hypoventilation takes place in NREM sleep [21]. Significant reductions in blood flow and metabolism are linked to NREM sleep, while overall metabolic rate and blood flow during REM similar to wakefulness [22].The alpha, beta, and gamma rhythms were attenuated in NREM sleep, while theta and delta rhythms rise with awakeness, followed by an increase in alpha and beta rhythms in REM sleep [31]. Detecting and understanding NREM and REM sleep stages can provide many ways to detect nervous system and body functional disorders. Except for the eye movement, middle ear ossicles, and respiratory system, the body is paralyzed during a Rapid Eye Movement (REM). Although the brain is less active during non-rapid eye movement (NREM), the body can still move. A sleep disorder like Narcolepsy is characterized by excessive daytime drowsiness and abnormal REM sleep regulation [23]. NREM sleep is related with the PARASOMNIAS or unusual sleep related behaviors, that take place while sleeping [23]. Because the body muscles are more active. Predicting NREM and REM sleep is deemed to be a novel process for the diagnosis of sleeping disorders. It also evaluates the
Abstract—Classification of sleep stages is one of the most important diagnostic approaches for a variety of sleep-related disorders. Electroencephalography (EEG) is regarded as a powerful tool for examining the association between neurological effects and sleep phases since it correctly identifies sleep-related neurological alterations. During Non-Rapid Eye Movement (NREM) and Rapid Eye Movement (REM) sleep phases, a number of nerve and bodily functions are affected and therefore hold an important role both in their functionalities. This work aims to classify NREM and REM sleep stages from sleep EEG data and present a noble SleepExplain model, an explainable NREM and REM sleep stage classification to explain its predictions. In this work, sleep stages were classified using Random Forest, XGBoost, and Gradient Boosting ensemble classification models. Overall, we obtained an accuracy of 92.54% (Random Forest), 94.25% (Gradient Boosting), and 94.30% (XGBoost). For explainable classification model, we utilized a game theoretic approach, SHAP (SHapley Addictive exPlanations) to offer a convincing explanation for the prediction. Index Terms—Machine Learning, Ensemble , XAI, electroencephalography, explainable, Sleep Stage
I. I NTRODUCTION Sleep is one of the basic biological activities that are required for relieving stress. It is the brain’s fundamental functions that are crucial for a person’s learning ability, performance, and physical activity [2], [3]. Understanding sleep quality in an easier manner is the most important and interesting topic in the field of neuroscience and sleep disorder diagnosis. Sleep stage scoring is the state of the art for analyzing human sleep [4]. The goal of sleep stage scoring is to find the stages of sleep that are important for identifying and treating sleep disorders [5], [6]. The continuous recording of several electrophysiological signals, termed as Polysomnographic (PSG) signals, is used for sleep stage scoring purposes [7]. The American Academy of Sleep Medicine (AASM) provides the most widely used
979-8-3503-4602-2/22/$31.00 ©2022 IEEE
1
body and nervous system characteristics during sleep. In the practical world, the sleep stage is performed through some manual processes. An expert measures and monitors the sleep scoring process manually. Notably, when it is a hand-operated procedure, errors can be coped up at any point. An automated strategy might be more reliable for regulating NREM and REM sleep. Several EEG and biosignal studies have been published to investigate the relationship between EEG biomarkers and neurologic prognosis in medical and healthcare [34]–[39]. The objective of this research is to develop a approach that can automatically classify NREM and REM sleep by explaining the prediction model using XAI. It will use multi-channel EEG signals to train machine learning models, and features will be retrieved from these signals. We intended to automate this sleep score technique using data from three EEG channels from three distinct sites (C4, O2, and F4). F4, C4, and O2 from the frontal, central, and occipital lobes were utilized for our investigation. The purpose of this study is to develop better models for forecasting NREM/REM sleep and explain those models using Explainable AI.
III. M ATERIAL A ND M ETHODOLOGY This study presents an efficient and reliable automatic sleep stage classification model of NREM and REM sleep classes on three-channel EEG signals. Alongside, we added an explainable AI (XAI) appraoch for better understanding the inside mechanism of the model. The explainable AI shows the most devoted features for the final outcome produced by the classification model. Notably, we focused on increasing the accuracy of the prediction model of the NREM and REM sleep stages and explain this model with SHAP based approach. A. Model Architecture Figure 1 depicts a general architecture of the proposed investigation. First, we acquired the EEG data from channels F4, C4, and O2 and extracted features from the raw signal using FFT (Fast Fourier Transform). To eliminate any 60 Hz AC noise from the neighboring electrical grid, the EEG signal was filtered. Then, from the noise-free signal, features are retrieved. Next, the model was trained using three ensemble machine learning models: Random Forest, Gradient Boosting, and XGBoost. Finally, we implemented an SHAP based explainable classification model to explain the outcomes produced by the prior model.
II. A SSOCIATED S TUDY In the majority of research, overnight recorded EEG signals are used to classify the sleep stages [27]. These proposed studies recommend applying a variety of feature reduction strategies in order to identify the relevant features. Various classification methods have been developed to classify the phases of sleep. None of them, however, classifies sleep as exclusively NREM and REM sleep. Satapathy et al. [24] proposed a method for identifying two stages of sleep, such as awake and sleep, using the ISRUCSleep [29] dataset. The overall accuracy for the awake and sleep stages was 91.67% and 93.8% respectively. Ellis et al. [25] provided an interpretable taxonomy of sleep stages, and their research classified sleep into five distinct phases. In this work, the PhysioNet Sleep-EDF [30] dataset was used. Santaji el al. [26] used another form of EEG data from sixty participants which were collected and preprocessed using an IIR (Infinite Impulse Response) in their research. In their study, three sleep classes (REM, NREM1, and NREM2) were classified with an overall accuracy of 95.36%. Santosh et al. [27] presented a machine-learning model with an ensemble approach by using ISRUC-Sleep [29] dataset. In their work, sleep stages have been classified as waking, NREM (as N1, N2, and N3), and REM sleep achieving an accuracy 91.10%. Shen et al. [28] proposed an enhanced machine learning model based on the essence of features applied to the ISRUC [30] dataset and classified sleep states into five stages, including waking, NREM (as N1, N2, and N3), and REM achieving an accuracy of 81.65%. Hussain et al. [31] used the neurological biomarkers of sleep phases to measure the delta wave power ratios (DAR, DTR, and DTABR). These measurements were evaluated by biomarkers due to their nature of decreasing during NREM sleep and increasing during REM sleep.
Fig. 1. Proposed framework of an explainable sleep stage classification model.
B. Dataset Description The dataset was collected from a sleep center named Haaglanden Medisch Centrum (HMC, The Netherlands) [32], [33]. The initial data files are combined with different types of signals. We selected EEG and F4 from frontal lob, C4 from central lob, and O2 from occipital lob. Frontal lob controls the voluntary movements of the body. The thoughts and analytical activities are controlled from central lob. Occipital lob is responsible for the eyesight. To characterize the sleep stages of a cerebral condition, the EEG is divided into five frequency subbands: delta wave (0Hz - 4Hz), theta wave (4Hz - 8Hz), alpha wave (8Hz - 12Hz), beta wave (12Hz - 30Hz), and gamma wave (30Hz). Our dataset contains 154 sleep recordings with 75 features. There are five classifications in the dataset, such as Wake, N1, N2, N3, REM. In the preprocessing
2
stap, all wake classified rows were removed and N1, N2, N3 classes were merged into a single NREM class. So we have received the final preprocessed dataset having two classes; NREM and REM. The final dataset has a total of 89096 rows, where rows containing NREM are 72631 and rows containing REM are 16465. The difference between the sample counts of NREM and REM classes is significant. So we implemented the SMOTE (Synthetic Minority Over-sampling Technique) [13] approach to balance our training dataset. The testing dataset was unchanged to find the original accuracy of our proposed model.
D. Explainability of the Proposed Model using SHAP It is crucial to be able to appropriately interpret the results of prediction models. It fosters the right level of user trust, offers suggestions for how to make a model better, and aids in comprehending the learning process that is being represented [40]. Explainable artificial intelligence (XAI) helps users to understand and believe the result produced by machine learning algorithms. Each feature is given a relevance value by SHAP (SHapley Additive exPlanations) for a specific prediction [1]. In this study, we took out the most important features and their contributions for a particular prediction applying SHAP value. It explains an outcome of a model by computing the contribution of each feature associated with it. The Shapely value of an outcome is produced following a linear model and measures how much each feature in that model contributes, either positively or negatively.
C. Classification Models We used three popular machine learning models Random Forest, Gradient Boosting and, XGBoost. The training dataset was balanced using SMOTE technique. The default values of performance parameters like n estimators and max depth did not produce a satisfactory result. So we had to tune the parameters of the algorithms. We used scikit-learn python library for training and tuning our models.
IV. R ESULT A ND D ISCUSSION In this section we will discuss the outcome of prediction performances. LaterSHAP detects the most influential features for XGBoost classifier. TABLE II P ERFORMANCE M ETRIC R ESULTS F OR M ODELS
TABLE I O PTIMUM PARAMETER VALUES Parameters max depth n estimators
RF 39 450
GB 12 1150
Evaluation Metric Accuracy Precision Recall Specificity F1 Score
XG 29 4010
Random forest is one of the most effective machine learning techniques for classification [14]. In this solution, we tuned this model with the n estimators covering the range of 3 to 500 and the max depth covering the range of 5 to 50. We figured out the best result when, n estimators = 450 and max depth = 39 having the accuracy of 92.52% Using an iterative process, boosting algorithms merge weak learners into a strong learner [15]. Gradient boosting is a regression approach that resembles boosting [16]. The accuracy with the default value of n estimators and best dept for Gradient Boosting Classifier was not quite acceptable. We tuned this model at the parameters, n estimators having the range of 5 to 1200 with a interval of 50 and best dept having the range of 3 to 30. The most acceptable result was acquired for n estimators = 1150 and best dept = 12. The accuracy with this tuning is 92.25%. Another ensemble model based on Gradient Boosting with a high degree of scalability is called XGBoost [17]. XGBoost constructs a loss function which is minimized if the objective function is expanded additively, similar to Gradient Boosting. The accuracy with the default value of n estimators and best dept for XGBClassifier Classifier was not also satisfactory. This model was tuned with the parameters, n estimators including the range of 500 to 5000 with an interval of 50 and the best dept including the range of 3 to 30. The topmost result of this model was attained for n estimators = 4010 and best dept = 29. It scored 94.3%.
RF 92.54% 87.13% 89.31% 89.31% 88.21%
GB 94.25% 91.25% 89.65% 89.51% 90.44%
XG 94.30% 91.33% 89.70% 89.70% 90.51%
A. Model Performance Result In Table II, the performance metric of three models are given. The best accuracy was achieved by XGBoost classifier and it was 94.30%, when other two classifiers Random Forest and Gradient Boost scored 92.54% and 94.25% respectively. XGBoost and Gradient Boosting scored almost same score. The differences is very tiny. In other Evaluation metrics, XGBoost performed very satisfactory. Gradient Boosting was also little bit behind from XGBoost. But Random Forest was far behind from other two classifiers. Individually, 92.54% is not a bad accuracy. But compare to other two classifiers, Random Forest was far behind. The ROC curve illustrates the trade-off between specificity and sensitivity. A better performance is shown by classifiers that provide curves that are closer to the top-left corner. The figure 2 is presenting the training-testing ROC graphs of Random Forest. In the top-left corner, differences between training and testing are slightly large than other two models. The AUROC for the training is 1.00, and for the testing the AUROC is 0.968. The training-testing ROC graph of Gradient Boosting is shown in The figure 3. The differences of the top-right corner of the graph was reduced The training and testing curves are closer than the Random Forest model. The accuracy is better The AUROC is 0.975 for the testing and 1.00 for the training.
3
Fig. 2. ROC Graph of Random Forest Model
Fig. 4. ROC of XGBoost Model
were naturally smooth. Though, we used SMOTE for the balancing the training data, the models performed well with the uncorrupted testing data. So we can assure that, the models are not overfitted.
Fig. 5. Confusion Matrix of XGBoost Model
Classification problem’s prediction outcomes are compiled in a confusion matrix. In the figure 5, the confusion matrix of XGBoost is shown. We can see the total prediction of NREM classes were 14460, where 14037 predictions were correct and only 423 predictions were wrong. Total REM classes wee 3350, 2766 were correct and 584 were wrong. We did not use SMOTE in testing data. That’s why, the imbalance is noticeable here. Though our models did perform well.
Fig. 3. ROC Graph of Gradient Boosting Model
The figure 4 is showing the training-testing ROC graphs of XGBoost. XGBoost performed better than other two models. The differences between XGBoost and Gradient Boosting is not not much. That’s why, the ROC curvs of this two models are significantly similar. The testing curve is more closer to the training graph than the other two graphs. The AUROC is 0.977 which close to the Gradient Boosting 3. There are no unusual ups and down in the ROC graphs of the three models. The Training and testing graphs were naturally smooth. There are no unusual ups and down in the ROC graphs of the three models. The Training and testing graphs
B. SHAP(XAI) Result The internal mechanism of a machine learning model is hard to understand and difficult to relate to the practical dataset and scenario. It is deemed to be a novel approach to use Explainable AI(XAI) to understand the inner mechanism of a model. In this study, we used SHAP for the better
4
understanding and feature dependencies. We applied SHAP in XGBoost classifier to determine the most significant features for a particular prediction. We considered the first 100 rows of the testing data.
XGBoost, and Gradient Boosting. Overall, Random Forest obtained 92.54% accuracy, XGBoost earned 94.30% accuracy, and Gradient Boosting also achieved 94.25% accuracy. We also applied SHAP in the XGBoost classifier to determine the most significant features for a particular prediction. Most of the traditional classifying approaches are for the five stage classification of sleep. Very few studies worked with XAI. So the inside mechanism of a model is not revealed. It remained a black box model. This study unleashed the model explainability with XAI and classified the two most significant sleep stages (NREM and REM) with better accuracy. ACKNOWLEDGMENT This research was supported by Islamic University of Technology Research Seed Grants (IUT RSG) (Ref: REASP/IUTRSG/2022/OL/07/012).
Fig. 6. Significant Features With SHAP Value
Fig 6 is showing the list of the most important features and their SHAP values for predictions. MedianF Beta F4 is the most dedicated feature. MeanF Alpha F4 is also scored almost same as Median Beta F4. Most of the features of the list scored +0.02. The rest of the features scored +0.41 altogether. Notably, the most of the significant features are from the frontal lob. The dedication of the occipital lob is less than other two lobs.
R EFERENCES [1] Scott M. Lundberg, Su-In Lee, “A Unified Approach to Interpreting Model Predictions,” 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. [2] Estrada, E and Nava, Patricia and Nazeran, Homayoun and Behbehani, Khosrow and Burk, J and Lucas, Itakura distance: A useful similarity measure between EEG and EOG signals in computer-aided classification of sleep. 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference, 2006 , pp. 1189–1192. [3] Li, Yi and Yingle, Fan and Gu, Li and Qinye, Tong, “Sleep stage classification based on EEG Hilbert-Huang transform,” in IEEE, 2009 4th IEEE Conference on Industrial Electronics and Applications, 1963, pp. 3676–3681. [4] Ebrahimi, Farideh and Mikaili, Mohammad and Estrada, Edson and Nazeran, Homavoun, “Assessment of itakura distance as a valuable feature for computer-aided classification of sleep stages,” 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2007, pp.3300–3303. [5] Mora, Antonio Miguel and Fernandes, Carlos M and Herrera, Luis Javier and Castillo, Pedro A and Merelo, Juan Julián and Rojas, Fernando and Rosa, Agostinho C, “TSleeping with ants, SVMs, multilayer perceptrons and SOMs,”2010 10th International Conference on Intelligent Systems Design and Applications, 2010, pp.126–131. [6] Vatankhah, Maryam and Akbarzadeh-T, Mohammad-R and Moghimi, Ali, “An intelligent system for diagnosing sleep stages using wavelet coefficients,” The 2010 International Joint Conference on Neural Networks (IJCNN), pp.1–5, 2010. [7] Aboalayon, Khald Ali I and Faezipour, Miad and Almuhammadi, Wafaa S and Moslehpour, Saeid, “Sleep stage classification using EEG signal analysis: a comprehensive survey and new investigation ,”2016, pp.272. [8] Shuyuan, Xiao and Bei, Wang and Jian, Zhang and Qunfeng, Zhang and Junzhong, Zou and Nakamura, Masatoshi, “Notice of Removal: An improved K-means clustering algorithm for sleep stages classification,”2015 54th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE), 2015, pp.1222–1227. [9] Lajnef, Tarek and Chaibi, Sahbi and Ruby, Perrine and Aguera, PierreEmmanuel and Eichenlaub, Jean-Baptiste and Samet, Mounir and Kachouri, Abdennaceur and Jerbi, Karim, “Learning machines and sleeping brains: automatic sleep stage classification using decision-tree multiclass support vector machines,”Journal of neuroscience methods, 2015, pp.94–105. [10] Güneş, Salih and Polat, Kemal and Yosunkaya, Şebnem, “Efficient sleep stage recognition system based on EEG signal using k-means clustering based feature weighting,”Expert Systems with Applications, 2010, pp.7922–7928. [11] Charbonnier, Sylvie and Zoubek, Lukas and Lesecq, Suzanne and Chapotot, Florian, “TSleeping with ants, SVMs, multilayer perceptrons and SOMs,”Computers in biology and medicine, 2011, pp.380–389.
Fig. 7. SHAP Value Beeswarm
The beeswarm plot(fig 7) is made to show a summary of the top features in a dataset and how they affect the model’s output in a way that is both information-dense and easy to understand. A single dot is used to indicate each instance of the explanation in each aspect of the figure 7. Here, MedianF Beta F4 is the most important feature on average. So from the both figure (fig 6 and fig 7), we can notice one thing and that is , F4 and C4 have a very significant role, where O2 has not such kind of contribution like other two lobes. The duration of REM sleep was less then the NREM. Eyesight and eye movement are controlled by the occipital lob. This is cause to reduce the supremacy of O2 lob in the model. V. C ONCLUSION In this study, NREM and REM sleep phases were classified using ensemble classification models from Random Forest,
5
[12] Berry, Richard B and Brooks, Rita and Gamaldo, Charlene E and Harding, Susan M and Marcus, C and Vaughn, Bradley V and others, “The AASM manual for the scoring of sleep and associated events,”Rules, Terminology and Technical Specifications, Darien, Illinois, American Academy of Sleep Medicine, 2012. [13] Nitesh V. Chawla,Kevin W. Bowyer ,Lawrence O. Hall,W. Philip Kegelmeyer, “SMOTE: Synthetic Minority Over-sampling Technique,”Journal of Artificial Intelligence Research 16 (2002) 321–357. [14] Leo Breiman. Random forests. Machine Learning, 2001, 45(1):5–32. [15] Robert E. Schapire Yoav Freund. A Short Introduction to Boosting. Journal of Japanese Society for Artificial Intelligence, 1999, 14(5):771 – 780. [16] Jerome H. Friedman. Greedy function approximation: a Gradient Boosting machine. The Annals of Statistics, 2001, 29(5):1189 – 1232. [17] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting sys- tem. In Proceedings of the 22Nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, New York, NY, USA, 2016. ACM, pages 785–794. [18] Institute of Medicine (US) Committee on Sleep Medicine and Research; Colten HR, Altevogt BM, editors. Sleep Disorders and Sleep Deprivation: An Unmet Public Health Problem. Washington (DC): National Academies Press (US); 2006. 2, Sleep Physiology. Available from: https://www.ncbi.nlm.nih.gov/books/NBK19956/. [19] Krieger J. Respiratory physiology: Breathing in normal subjects. In: Kryger M, Roth T, Dement WC, editors. Principles and Practice of Sleep Medicine. 4th ed. Philadelphia: Elsevier Saunders; 2000. pp. 229–241. [20] Simon PM, Landry SH, Leifer JC. Respiratory control during sleep. In: Lee-Chiong TK, Sateia MJ, Carskadon MA, editors. Sleep Medicine. Philadelphia: Hanley and Belfus; 2002. pp. 41–51. [21] NLM (National Library of Medicine), NIH (National Institutes of Health). Medline Plus Online Medical Dictionary. [accessed February 6, 2006]. [22] Madsen PL, Schmidt JF, Wildschiodtz G, Friberg L, Holm S, Vorstrup S, Lassen NA. Cerebral O2 metabolism and cerebral blood flow in humans during deep and rapid-eye-movement sleep. Journal of Applied Physiology. 1991b;70(6):2597–2601. [23] Kathryn Lovell, PhD, and Christine Liszewski, MD, Normal Sleep Patterns and Sleep Disorders [24] Satapathy, S.K., Loganathan, D. A Study of Human Sleep Stage Classification Based on Dual ChannelSatapathy, S.K., Loganathan, D. A Study of Human Sleep Stage Classification Based on Dual Channels of EEG Signal Using Machine Learning Techniques. SN COMPUT. SCI. 2, 157 (2021). [25] Ellis CA, Zhang R, Carbajal DA, Miller RL, Calhoun VD, Wang MD. Explainable Sleep Stage Classification with Multimodal Electrophysiology Time-series. Annu Int Conf IEEE Eng Med Biol Soc. 2021 Nov;2021:2363-2366. doi: 10.1109/EMBC46164.2021.9630506. PMID: 34891757. [26] Santaji, S., Santaji, S., Desai, V. Automatic sleep stage classification with reduced epoch of EEG. Evol. Intel. 15, 2239–2246 (2022) [27] Santosh Kumar Satapathy, Akash Kumar Bhoi, D. Loganathan, Bidita Khandelwal, Paolo Barsocchi, Machine learning with ensemble stacking model for automated sleep staging using dual-channel EEG signal, Biomedical Signal Processing and Control, Volume 69, 2021, 102898, ISSN 1746-8094, https://doi.org/10.1016/j.bspc.2021.102898. [28] Shen H, Ran F, Xu M, Guez A, Li A, Guo A. An Automatic Sleep Stage Classification Algorithm Using Improved Model Based Essence Features. Sensors (Basel). 2020 Aug 19;20(17):4677. doi: 10.3390/s20174677. PMID: 32825024; PMCID: PMC7506989. [29] Khalighi Sirvan, Teresa Sousa, José Moutinho Santos, and Urbano Nunes. “ISRUC-Sleep: A comprehensive public dataset for sleep researchers.“Computer methods and programs in biomedicine 124 (2016): 180-192. [30] B Kemp, AH Zwinderman, B Tuk, HAC Kamphuisen, JJL Oberyé. Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the EEG. IEEE-BME 47(9):1185-1194 (2000). [31] Hussain, I.; Hossain, M.A.;Jany, R.; Bari, M.A.; Uddin, M.; Kamal, A.R.M.; Ku, Y.; Kim, J.-S. Quantitative Evaluation of EEGBiomarkers for Prediction of Sleep Stages. Sensors, 22, 3079 (2022). https://doi.org/10.3390/s22083079 [32] Alvarez-Estevez, D.; Rijsman, R.M. Haaglanden Medisch Centrum Sleep Staging Database (Version 1.0.1); PhysioNet, 2021; Availableonline: https://physionet.org/ (accessed on 22 March 2022).
[33] Alvarez-Estevez, D.; Rijsman, R.M. Inter-Database Validation of a Deep Learning Approach for Automatic Sleep Scoring. PLoSONE 2021,16, e0256111. [CrossRef] [34] Hussain, I. and S. J. Park. ”Big-ecg: Cardiographic predictive cyberphysical system for stroke management.” IEEE Access 9 (2021): 123146-64. [35] Hussain, I. and S. J. Park. ”Healthsos: Real-time health monitoring system for stroke prognostics.” IEEE Access 8 (2020): 213574-86. [36] Hussain, I., M. A. Hossain and S.-J. Park. ”A healthcare digital twin for diagnosis of stroke.” Presented at 2021 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON), 2021. 2022. [37] Hussain, I. and S.-J. Park. ”Quantitative evaluation of task-induced neurological outcome after stroke.” Brain Sciences 11 (2021): 900. [38] Hussain, I., S. Young, C. H. Kim, H. C. M. Benjamin and S. J. Park. ”Quantifying physiological biomarkers of a microwave brain stimulation device.” Sensors 21 (2021): 1896. [39] Hussain, I.; Young, S.; Park, S.-J. Driving-Induced Neurological Biomarkers in an Advanced Driver-Assistance System. Sensors 2021, 21, 6985. [40] M. K. Sumon, M. H. Ashmafee, M. R. Islam, and A. R. Mostofa Kamal, “Explainable NLQ-based visual interactive system: Challenges and objectives,” Proceedings of the 2nd International Conference on Computing Advancements, 2022.
6