A Novel Approach for Sleep Arousal Disorder Detection Based on the Interaction of Physiological Signals and Metaheuristic Learning

The vast majority of sleep disturbances are caused by various types of sleep arousal. To diagnose sleep disorders and prevent health problems such as cardiovascular disease and cognitive impairment, sleep arousals must be accurately detected. Consequently, sleep specialists must spend considerable time and effort analyzing polysomnography (PSG) recordings to determine the level of arousal during sleep. The development of an automated sleep arousal detection system based on PSG would considerably benefit clinicians. We quantify the EEG-ECG by using Lyapunov exponents, fractals, and wavelet transforms to identify sleep stages and arousal disorders. In this paper, an efficient hybrid-learning method is introduced for the first time to detect and assess arousal incidents. Modified drone squadron optimization (mDSO) algorithm is used to optimize the support vector machine (SVM) with radial basis function (RBF) kernel. EEG-ECG signals are preprocessed samples from the SHHS sleep dataset and the PhysioBank challenge 2018. In comparison to other traditional methods for identifying sleep disorders, our physiological signals correlation innovation is much better than similar approaches. Based on the proposed model, the average error rate was less than 2%–7%, respectively, for two-class and four-class issues. Additionally, the proper classification of the five sleep stages is determined to be accurate 92.3% of the time. In clinical trials of sleep disorders, the hybrid-learning model technique based on EEG-ECG signal correlation features is effective in detecting arousals.


Introduction
An arousal during sleep is often thought of as a sleep disorder caused by fragmented sleep [1]. Changes in mobility, motor consciousness, and response readiness are caused by arousal-induced changes in sleep because they raise blood pressure and heart rate [2]. Multiple system atrophy (MSA), Lewy body disease (LBD), and Alzheimer's disease (AD) can all be detected early by studying physiological sleep signals [3]. Te precise recognition of sleep stages and patterns can be crucial in reducing the progression of neurological diseases in such situations [4]. According to the American Academy of Sleep Medicine (AASM), monitoring physiological markers can detect rapid changes in sleep stages [5]. Consequently, classifcation of sleep stages, investigation of heart rate variability, REM sleep without atonia (RSWA), polysomnography (PSG), electrooculography (EOG), electroencephalography (EEG), electrocardiography (ECG), and electromyography (EMG) have received less attention in detecting arousal (EMG). To make easier the recognition of neurocognitive disorders, EEG is a cost-efective and promising test [6]. Trough activation of the autonomic nervous system (ANS), cortical arousals are linked to fuctuations in heart rate and blood pressure [7][8][9]. Neurologists recognize arousals as abrupt changes in the EEG frequency [10], so they prefer to use EEG signal analysis to detect them. Although recording EEG data during sleep testing can be cumbersome for most individuals, most arousal episodes are associated with ANS activity. Sleep stage changes and other related problems are commonly accompanied by increased heart rate variability [11][12][13]. Arousals, also called cortical and autonomic activation, can occur spontaneously or as a result of respiratory dysfunction during sleep [14]. Physiological markers such as EEG and ECG and ANS activation during sleep arousal have been studied previously [15].
Nonlinear dynamical systems (NDSs) are thought to possess key properties based on their metric disparities between events [16]. Due to the complex patterns of such metric discrepancies, quantifying such diferences is a diffcult task for recognizing a correlation. Physiological sleep signals are assessed and managed nightly by neurologists [17], which is an empirical, inefcient, and time-consuming process [18]. Terefore, automatic interpretation and recognition are desirable.
Some previous research studies attempted to undertake automatic arousal detection based on physiological signals, with varying degrees of success. Attempts have been made to establish a link between physiological signals and sleep stage shifts [12,13]. Te study [14] has examined the relationship between arousal and the heart rate response to sleep stage changes. Despite the fact that the standard approach implies that at least one EEG channel is benefcial, some contributors have suggested that alternative physiological signals be used [19]. Te difculties of using a semiautomated system to detect and measure arousal have been highlighted in ECG signals [20]. Pillar et al. described an automated detection technique for separating AASM-defned arousals from sleep using peripheral arterial tonometry and a mobile device [21]. Warrick and Nabhan Homsi [22] proposed identifying sleep arousal using PSG and recurrent neural networks (RNNs). Deep learning (DL) has recently been widely employed to analyze EEG and ECG signals, including convolutional neural networks (CNNs) for sleep stage scoring [23][24][25][26], and an end-to-end (e2e) deep learning method based on multiphysiological inputs [27]. Tsinalis et al. [28] used timefrequency analysis and stacked sparse auto-encoders (AEs) to create an autonomous sleep stage grading system. Macias et al. [29] used the wavelet transform (WT) to classify physiological sleep data and then used neural network (NN) to detect alertness. Zhou et al. [30] reported a developed ensemble deep learning (EDL) structure that retains temporal associations of multimodal physiological inputs using a positional embedding-based multihead attention architecture. Ugur and Erdamar [31] used an ensemble bagged tree to assess quantitative sleep EEG synchronization for automatic arousals detection and achieved 99.5% accuracy. From PSG recordings, Karimi and Asl [32] proposed an automatic identifcation approach for nonapneic sleep arousal zones. Tey developed ensemble techniques and concentrated on feature subset selection and consensus approaches.
Nonlinear analysis of the EEG signal has long been recognized as providing information beyond that provided by regular EEG metrics [31]. Pathological events and normal cognition have both been shown to have diferent levels of brain complexity [32]. Te chaotic phenomena are one of the primary causes of EEG signal complexity, and chaotic dynamics are characterized by sensitive dependency on initial conditions. Moreover, a Lyapunov exponent [33] is a method for describing the chaotic nature of EEG signals. In addition, some research has shown that the R segment of ECG signals can exhibit fractal behavior [34]. Te classic and efective techniques for constructing discriminating features from physiological signals is single spectrum entropy analysis [35], entropy [36], complexity [37], Lyapunov exponent [38], fractal dimension [39], and wavelet transform [40].
Comparing existing studies in an unbiased and fair manner is very difcult because there are so many approaches to study the same subject [41,42]. For datasets that are not publicly available, physiological signals and evaluation performance metrics difer. A major problem is the lack of standards for detecting sleep arousal, which is what creates this issue.

Motivation and Contribution
Te frequent awakening from sleep, in addition to disrupted and fragmented sleep, is a precursor to somnipathy, which can lead to a variety of future health problems. Problems with sleep arousal have been linked to obesity, depression, heart disease, and diabetes. In order to improve public health, we need a deeper understanding of arousal neurophysiology.
It is possible to diagnose and treat sleep issues if excessive arousals are detected early. Early recognition of a patient's health condition can prevent changes in blood pressure and heart disease. Sleep arousal is often determined by PSG recordings. Furthermore, sleep-awakening events are subjectively assessed and vary greatly among professionals. In order to determine the level of arousal during sleep, sleep experts must examine PSG recordings. Clinicians would greatly beneft from an automated PSG-based sleep arousal detection system. Two major obstacles must be overcome (1) understanding arousal disorders, which can cause disease, and the detrimental efects of excessive arousal on regular sleep. (2) Te neurologist's outcome must be consistent if the model creates sleep stages and detects arousal disorders.
Taking care of these issues may have a substantial impact on the treatment process, and the individual may be cured of neurological ailments in the future. Tus, a precise approach for assessing arousals and sleep stages must be developed to monitor the diagnosis. Experts can signifcantly beneft from improving decision making through meta-learning and optimization methods in this context [43].
In this study, Lyapunov exponents, wavelet transforms, and fractals are used to quantify sleep and arousal stages through feature extraction from EEG-ECG signals. For the frst time, an efcient hybrid-learning method is presented in this paper to detect and assess arousal incidents. In this novel model, a modifed drone squadron optimization (mDSO) algorithm [44] is used to optimize support vector machines (SVMs) with radial basis functions (RBFs). As compared to other traditional methods for identifying sleep disorders, our physiological signals correlation innovation is much more efective. According to the proposed model, there was an average error rate of less than 2% for two-class issues and 7% for four-class issues, respectively.
A key contribution of this study is (a) the development of correlations between physiological signals, (b) the development of feature extraction methods, and (c) the design of the optimal classifcation scheme for the diagnosis of arousals, sleep stages, and sleep disorders.
We also present evidence supporting the utility of EEG and ECG signals with correlated brain and autonomic activity efects. Te analyses of ECGs and EEGs enabled a more detailed understanding of arousal and sleep stages. An EEG-ECG system can also be used to quantify psychological sleep events such as arousal disorders and sleep stage changes as shown in Figure 1. It is common for physiological signals of arousal to change abruptly from deep to light sleep, which is why sleep stages were also identifed. Due to this, the authors decided to classify sleep steps as well as determine arousal levels.
Tis paper is organized as follows: Section 3 introduces the learning structure. Section 4 describes the method proposed for identifying abrupt changes and sleep stage. In Section 5, statistical analyses and results are discussed. Finally, conclusions are drawn in Section 6.

Learning Structure
To optimize the classifcation approach based on the learning process, this section discusses the components of the learning structure.

Classifcation Learning.
Let us assume that we need to obtain a determination function f with the characteristic f (x j ) � y j , ∀j [45].
A discriminating hyper-plane typically does not exist. Te slack variables ζ j are ofered to provide for the event of examples violating (1): Moreover, w is a weight vector, x is the input vector, and b is bias. As a result, we have Te support vector (SV) procedure to minimize the assured uncertainty bound includes Te soft margin equation is made by multiplying ζ by a hyperparameter c and adding two terms with l training points. Te minimization procedure is conducted based on mentioned constraints in (2)-(3). Importing Lagrange multipliers α j and applying the Kuhn-Tucker theory of the optimization approach, the answer can be determined w � l j�1 y i α j x j . (5) Tis condition in (3) must be precisely met by the relevant cases (x j and y j ) in order for α j to be nonzero. Tese x j are known as "support vectors" and the rest of the training examples are unnecessary. Teir exclusion from the expansion (5) is due to the fact that (3) is immediately satisfed (with ζ j � 0). Calculating α j 's coefcients entails solving the quadratic programming issue as shown. Tus, the maximization of the following equation can be applied with regard to (7): Based on the dot product's linearity, the decision function can also be constructed as follows: where, the sgn corresponds to the sign function or signum function. A set of input vectors x 1 ,. . ., x l can be nonlinearly translated into a high-dimensional feature space to provide signifcantly more general decision surfaces. Te decisionmaking function is As a result, kernel RBF can be displayed as the following for support vectors: 3.2. Drone Squadron Optimization. Te drone squadron optimization (DSO) is comprised of two components: semiautomated drones that conduct environmental surveys and a command center that processes data and updates the drones' frmware system as needed. Te self-adaptive component of DSO used in this work is the perturbation/ motion method used to get the target coordinates [44]. Te command center devised this method of customizing the DSO to the search perspective during the global optimization phase.
Drones' capacity to be changed or modifed in both hardware and software is one of its most distinguishing features (the frmware). As a result, researchers may quickly include new processes into the algorithm because these devices are controlled by software (frmware). Te command center serves as a single point of contact for issuing orders and overseeing assigned responsibilities. Internal decisions Computational Intelligence and Neuroscience are made based on the system's processing of inputs and data. Te command center can update the drone software at any time, adapting a team's behavior to the scenario dynamically.
Drone squadrons and command centers make up the DSO algorithm. Te command center accomplishes two goals by utilizing data collected by/from drones. To keep some control over the search process while also developing new drone-fying technology. Drones are not the answer because they are programmed to fy to a specifc location (the actual solution). It consists of the methods and parameters used by teams when doing landscape searches.
Tere must be a departure point in order to avoid limiting the search area towards the origin assuming that the ofset value is zero, the departure coordinates still have noise app lied to them, which causes a neighborhood search. If a frmware update condition is reached (for example, the number of iterations), the command center substitutes variations of the w best frmware for the w worst frmware. For w � 1, this indicates that the team with the lowest cumulative rank will obtain a frmware version (random subtree replacement) of the best team.

Proposed Methodology
Feature extraction, feature selection, and optimized classifcation are steps of the proposed strategy ( Figure 2).

Preprocessing.
We must window nonstationary physiological signals in order to conduct the essential analysis. Each window overlaps the previous one to preserve the information's nature. Due to the fact that EEG and ECG signals vary in length across states, we separated them into equal segments based on the sampling frequency. In other words, we used the windowing technique, which divides the signal into equal time segments. By considering a 20%-50% overlap between two consecutive windows, the length of each window must be calculated in such a way that appropriate features may be extracted based on the established step-times [46]. Te rationale for selecting the signal length and the percentage of overlap between the consecutive signals will be discussed in detail in the section on experimental results.
By generating deep sleep in the third stage and allowing it to take its course naturally in the ffth stage, band separation is accomplished utilizing EEG signals. Te third stage is defned by the presence of slow, deep brain waves termed waves. We created low-pass flters with passing frequencies of 3 Hz or fewer to diferentiate EEG signals from waves. Individuals' attentiveness decreases during the deep sleep stage, which means that nearby noises or activities may go unnoticed. Tis stage serves as a transition between awake and deep sleep. Dreaming typically happens at step 5, sometimes referred to as rapid eye movement (REM). Deep sleep, which is frequently referred to as partial sleep, is defned by increased brain activity, eye movement, and rapid breathing. Using the linear regression algorithm (LRA), we can similarly distinguish cardiac complexes from signals [47].

Feature Extraction and Selection.
Te feature extraction level includes three descriptors as follows:

Lyapunov's Exponent.
In phase space, we can determine the convergence or average exponential divergence of surrounding trajectories by using Lyapunov exponents [48]. Generally, chaotic systems have a positive Lyapunov ex- Signal Windowing state vector p of a dynamic system. As the system progresses, we measure the evolution of the distance d between the adjacent phase points p 1 and p 2 in the dimensional space and generate trajectories [p 1 (t), p 2 (t)]: where | ε → (t)| is a very small and positive value. When chaotic dynamics are present in system Using the average velocity of the trajectories as a starting point, we can determine the exponential divergence of the average velocity: In exact terms Te following matrix shows the extracted features from overlapped-windowed EEG/ECG signals. We form two independent feature vector matrixes for each physiological signal. Following is an example using the Lyapunov exponent and a sample feature extraction vector from each windowed signal based on three descriptors. (A) Correlation Dimension. In the fractal theory of a signal, the correlation dimension (denoted by v) is a measure of the dimensions of space occupied by a set of random points, which is known as the fractal dimension. In the frst dimension, the distance between every two-point pairs in feature space is obtained based on the phase environment. In the second step, the summation of C (N and r) and D 2 parameters are estimated according to  Computational Intelligence and Neuroscience where ||.|| represents the Euclidean distance and r is the radius of the V (j) neighborhood. Also, N represents the number of points, θ is a step function, and fnally, V is the phase-space vector, which is determined according to the following In this equation, m is the embedding dimension, and t is the time delay.
(B) Fractal Dimension. Te degree of complexity, irregularity, and chaos of the signal is determined by the fractal dimension.

Wavelet Descriptor.
Te discrete wavelet transform (DWT) was created because of the need for multiscale feature representation. DWT refects physiological information in a variety of ways, depending on the size or scale. Selecting decomposition levels based on the frequency component that dominates the EEG and ECG data. A high degree of correlation between the levels of the signal sections and the frequency data are essential for categorizing signals. For our experiment, the study uses a fve-level decomposition structure. As a result, physiological signals are divided into D1-D5 (detail coefcients) specialties and an A5 (approximation coefcients) approach. Wavelet types are typically investigated until the most appropriate one is found. A Daubechies wavelet is more suited for detecting signal changes because of its second-order (db2) smoothing characteristics. Terefore, db2 was used to compute the wavelet coefcients for this study.

Feature Selection.
Selecting features based on generic flters, we can use the shared wrapper feature selection as a recursive feature eradication for clusters, including the knearest appropriate neighbor and support vector machine. To do so, we allocate a score called Sc i to each group, as explained by the following. Groups are then ranked relative to these assigned scores, and top-ranking groups are identifed as the primary outputs.
where δ i+ and ε i+ are the mean and standard deviation of features associated with abrupt changes in the sleep stage, and δ i− and ε i− are the mean and standard deviation normal change patterns in the sleep stage, respectively.

Optimized Learning.
A hybrid classifcation model based on the SVM learning repository is used in the frst part. Te initialization of the c (i.e., is the inverse of the standard deviation parameter in the kernel) and C (a soft margin parameter) in the classifcation kernel is randomly generated for each classifer. Te speed of change of each of these qualities is diferent. Te modifed DSO method is formed with a selected structure based on the least amount of error in the training data. Tis method accelerates and fne-tunes convergence toward the optimum response. Tis technique is repeated fve times to evaluate the classifcation fusion of the pool structure, the average accuracy, and the best matching parameters. Te network accuracy is a key factor in determining an appropriate match. Calculating the optimal values of the RBF kernel helps to search for the modifed DSO algorithm for global optimizations. Scaling, remodeling, and connection methods are used in the traditional DSO strategy, as well as selecting lowerquality solutions when stagnation occurs and frmware updates. Te initial unpredictability of the procedure, as well as the occurrence of desperate portions in the search region, are, however, faws. To reduce unpredictability and avoid desperate sections of the search region, we employ k-nearest neighbor (k-NN) clustering technique. Te DSO algorithm's initialization challenge is solved, enabling the fltering of incorrect initialization and, as a result, an improvement in the fnal response. Te K-NN classifer signifcantly aided in the determination of optimal starting values while using the standard DSO technique. K can be unlimited in the k-NN algorithm, which indicates that all points in the data are used to make the prediction, although their efect is inversely proportional to their Euclidean distances from the initial optimal targets. Te following equation illustrates the predictor, where is the index of the k-NN Te traditional DSO technique was repeated numerous times with varying values to create improper and appropriate starting points for use in the k-NN learning process, which were labeled 1 and 2, respectively, in proportion to the ultimate accuracy gained by establishing the best RBF kernel parameters.

Dataset
. Te proposed technique was tested using the Sleep Heart Health Study (SHHS) and PhysioBank databases, which were part of the Challenge 2018 project. Te SHHS consists of arousal events and various transitions from deep sleep to light sleep. Te National Heart, Lung, and Blood Institute has launched the SHHS, a multipurpose cohort study, to determine cardiovascular and other respiratory outcomes in the absence of any sleep disturbance. Between November 1, 1995, and January 31, 1998, 6,441 men and women over the age of 40 were registered for the SHHS visit.
As shown in Figure 3, physiological signals such as ECG, EEG, ABD, SaO2, chest, and AIRFLOW were annotated by neurological experts from numerous participants in both data and distinct periods. Wake state (W), sleep stage 1 (S1), sleep stage 2 (S2), sleep stage 3 (S3), sleep stage 4 (S4), and REM sleep were all given abbreviations. In both databases, the events captured in the crucial EEG signal are in the form of arousal. Te cardiac signal's duration is measured in minutes and seconds, just like the EEG signal. During the incident, the percentage decrease in the amount of dissolved oxygen in the blood was also recorded. Te lowest levels of dissolved oxygen in the blood were seen in general. Each record comprised all of the comments from the specialist's manual registration fles that could be identifed based on the type of occurrence or sleep stage.

Setting.
To run the algorithms in Matlab 2019b simulation environment, a computer system with 8 GB of RAM and a 64-bit operating system was required for Intel (R), Core (TM), and Core i7 CPU. Te synchronization of both signals acquired from the volunteers' sleep phases, which included six sleep steps, was considered equal to identify the sleep stages due to the signifcant duration of specifc signals obtained from the participants.
Te arousal events classifcation problem is a two-class task for EEG, ECG, and EEG-ECG signals. All ECG and EEG signals were preprocessed into discrete windows (at 1-second intervals) and deep sleep bands, with levels in the ECG signal that were higher than the EEG signal, however proportionate to it. It was extracted at the feature extraction stage along with Katz and Higuchi features (extraction from all separated signal frames) the following features from EEG and ECG signals: correlation, entropy, energy, skewness, standard deviation, mean, power ratio, spectral power, zero crossing, and reverse of diferentiation (with vector features ranging in length from 47 to 60).

Assessments.
In the assessment step, we partition data into test and training data using K-fold cross-validation (CV). In all, the signals were divided into four groups. Te semiconscious and awakening periods (also known as the EEG signal with beta and alpha sub-bands) as well as their homologous time intervals in the ECG signal set were deleted from brain datasets. Te frst signifes a rapid transition from light to deep sleep (Class 1). Te second class denotes deep sleep (Class 2), the third class denotes low-intensity sudden deep-sleep type shifts (Class 3), and the fourth class denotes rapid deep-sleep style changes (Class 4). In the frst report of outcomes, we compare learning based on particle swarm optimization (PSO), genetic algorithm (GA), and DSO algorithms with mDSO algorithm in improving the SVM classifer. As shown in Tables 1 and 2, we compared the performance of the SVM classifer in classes two and four using the mDSO optimizer to alternative optimization strategies. As demonstrated in Table 1, the binary classifcation indicated whether or not there was a sudden transition from deep to light sleep.
Te results of the sleep stages classifcation are depicted as confusion matrices in Figures 4-6. Four folds are available in these fgures, with two maximum and two minimum results, and the CV is preferred at 10.
Te matrices shown in Figures 4 and 5 are the results of the sleep stage classifcation on datasets 1 and 2. Te classifcation outcomes of sleep stages are displayed as confusion matrices in Figures 4-6. Te outputs of applying the approach to both datasets are also shown in Figure 6.
HRV analysis has been employed for decades to assess changes in the autonomic nervous system (ANS), which is the balance between the parasympathetic and sympathetic nervous systems. A change in the heart rate is one of the main indicators of cerebral cortex arousal [49][50][51][52][53][54][55]. Experts have identifed a stronger correlation between clinical outcomes and heart rate analysis, which sheds light on the systemic efects of arousal and sleep phases. Te fndings in this fgure represent the results of using the suggested method to both datasets, which were evaluated by concatenating the extracted features from both signals. Only the experiment's results are shown in these fgures. To prevent overftting in learning, the data are also shufed. Te suggested method's sleep stage classifcation fndings are included in the confusion matrix. Te primary purpose of presenting repeats is to demonstrate the degree of the apparent variance among the acquired data by computing the standard deviation (STD), which is used to verify the method's generalizability and robustness in diagnosing arousal. Te frst dataset's distribution accuracy is higher, while the divergence in the second dataset's results is minor. Te results of the classifcation of arousal and nonarousal events for EEG, ECG, and EEG-ECG signals are shown in Figure 7. Arousal may be identifed with a greater precision in the presence of both signals and due to their correlation, i.e., the concatenation of information from both signals.
Te comparison is based on the algorithm being applied to both datasets. Te appearance of arousal events is also scored by a neurologist. Te accuracy, AUC, and recognition rate (RR) were assessed, and the cross-validation approach was investigated by setting CV � 10. Tables 3-5 illustrate the accuracy, AUC, and RR of feature selections obtained from Computational Intelligence and Neuroscience     Table 5, all processes of feature extraction, feature selection, and classifcation were analyzed employing an improved SVM under various situations. Tese tables show how to use classifcation and apply features to EEG or ECG signals in a variety of situations.

Discussion
Tree experiments are provided in Figure 8 to assess the infuence of concatenated features of (a) ECG, (b) EEG, and (c) EEG-ECG signals. In the distance of 30-50% of all features, the features are selected because a proper discriminated efect in the classifcation.
For both the EEG-ECG signals, the FS, DWT, and LE descriptors retrieved a total of 112 features, of which 34-40 concatenated features were chosen as the best subset features in Figure 8(c). Te efective features, depending on the signals, have a higher impact on the fnal classifcation accuracy and are critical to the arousal diagnosis in all three trials shown in Figure 8. For identifying sleep disorders, the same is true, with 30-50 percent of the selected criteria yielding the highest classifcation accuracy. Diferent types of data, such as test and validation signals, were analyzed. As illustrated in Figure 9, the convergent level of the error is calculated using the loss function over a fnite number of epochs. Te mDSO procedures illustrated convergence of optimizing the SVM classifer is rapid and reliable for both test and validation data, as seen in this image. As a result, we investigate SVM in order to achieve optimal classifcation conditions by altering data.
Te Jaccard and average degree (AD) benchmarks are shown in Figure 10 to measure the quality of extracted features from all three FS, DWT, and LE descriptors, which are used to test the feature integration. 0.0% 0 Figure 4: Te confusion matrix is depicted in this table, which includes the frst (best) and second (worst) rows of the two-fold from the 10folds of EEG and ECG signals in the frst dataset.
Computational Intelligence and Neuroscience 9 Te fndings of its two-stage repeat, as well as for both EEG and ECG physiological signal models, show that it has a good performance in identifying sleep stage alterations. Te most efective classes for identifying transitions in sleep stages (Class 1 to Class 4) for EEG-ECG data are presented in this comparison (Class 1 to Class 4). What was thought to be a 5-repeat accuracy criteria experiment in arousal disorder detection turned out to be inconsequential, with modest variations recorded. Te outputs from the usage of the hybrid descriptor and mDSO procedure between EEG/ECG signals were more signifcant with the neurologistss' opinions, as shown in Figures 11 and 12. Te results are compared as a bar-plot based on combined features in Figure 11 for the sleep stage classifcation.
Two diferent types of data are used to assess the method's performance: EEG and ECG signals. Te advantages of the mDSO method over other methods for increasing learning include its confdence in achieving the optimal response and its rapid convergence.
Unlike prior studies that concentrated exclusively on arousal, the current research separated the signifcant stages of sleep to assess the intensity and intervals of arousal events. Other methods have either used cardiac signals or evaluated only EEG signals, with no consideration of the factors underlying their association in the response. Accuracy levels higher than 90% for diferent classes of sleep stages, and thus identifying arousal events using traditional algorithms, require the extraction of precise patterns of features.
Some algorithms in this feld have been successful in classifying arousal episodes; nevertheless, issues such as generalizability and computational complexity for big data still exist [25,26,[56][57][58][59][60]. Tere has not been a study of the simultaneous recognition of sleep stages, arousal detection, and the correlation between EEG and ECG signals.  Figure 5: Te confusion matrix is depicted in this table, which includes the frst (best) and second (worst) rows of the two-fold from the 10folds of EEG and ECG signals in the second dataset.
analyses similar arousal detection approaches and compares the performance of the suggested method-based measures such as accuracy. We used multiple signal channels and employed the method on a large number of similar datasets, whereas analyzing features and connecting signals using singlechannel signals is a challenge that must be addressed to ensure precision, both in the detection of sleep stages and in arousal events of varying intensities [23,31]. To extract useful features, we employed a hybrid approach. Deep learning approaches [23][24][25][26]61] may be able to extract automated separable features for the categorization. Nonetheless, extensive data and prolonged training complicate the signal analysis, i.e., the identifcation of hyperparameters and variables of training, high susceptibility to the uncertainty efect, and similar items in variation analysis parameters. What is referred to as arousal detection in EEG and even ECG analyzing may yield interpretations other than arousal based on signals such as PSG [22,32,60]. As a result, the proposed technique has a signifcant advantage over similar methods that uses single nonstationary and noise PSG or EEG signals [25,26,59] to detect sleep stages. Tus, by identifying combined features deduced from signals with richer EEG and ECG information for accurate arousal analysis.
Study limitations included a limited number of patients, complex physiological signals, and inaccurate labeling when it came to identifying arousal. Moreover, there is a lack of confrmation of a relationship between physiological signals during arousals, lack of an indicator boundary between arousals, and insufcient substrate to the estimate arousal intensity. Noise, including motion noises, is one of the biggest challenges in arousal and sleep stages detection. Since the EEG signal also includes artifacts such as  Figure 7: For ECG (frst raw), EEG (second raw), and EEG-ECG (third raw) signals, the results of a confusion matric (best three-folds) for the classifcation of arousal and nonarousal events are shown.    Computational Intelligence and Neuroscience respiration signals and movement signals, detecting the signal's noise component can be challenging. A second type of noise is produced when the power frequency of the recording equipment coincides with that of the well-known noise [62]. Even though we extracted features using several methods and windowed the signal, it is proposed that the Iteration (d) Figure 9: Te comparison of mDSO convergence based on the cost function error calculation. Te mDSO is a speedy algorithm that can quickly reach the ideal value. As a result, we employ SVM for reaching optimal classifcation conditions by adjusting data. (a, b) Test data and (c, d) validation data applied to the improved classifer.  Computational Intelligence and Neuroscience method, in addition to fltering the signal based on previous suggestions regarding the signal decomposition using efcient methods such as wavelet [63], also prevents information from being lost by physiological signals by preventing information from being lost. In the future, authors will use graph convolutional network (GCN) [64] and improved transfer learning networks [65] to determine arousal events from EEG to ECG data.

Conclusion
Tis paper provided an efective approach for extracting relevant characteristics from EEG to ECG data and categorizing them using optimal learning. In addition, we develop a metaheuristic learning model based on physiological sleep signals to detect arousal disorders. Tis study also intends to defne distinct sleep stages as well as rapid transitions from deep to light sleep stages, which is a short-term abnormality in EEG-ECG signals. Nonetheless, it ofers critical information on the likelihood of developing Alzheimer's, Parkinson's, or MSA illnesses. Tere is strong evidence of a relationship between these problems and the diseases indicated above. Te arousal events were described using an optimal learning model based on the mDSO and SVM. Using concatenated features, we were able to develop an efcient discriminative approach that analyzed physiological signals using three descriptors. Future research could look into the relationship between ECG and EEG signals as a way to prevent numerous diseases caused by sleep disturbances.

Data Availability
All the codes are available from the corresponding author upon request. Te data that support the fndings of this study were derived from the resources available in the public domain (https://sleepdata.org/datasets/shhs) and (https:// archive.physionet.org/physiobank/database/challenge/2018).