Ea sensors Review Electroencephalography Signal Processing: A Comprehensive Review and Analysis of Methods and Techniques ‘Ahmad Chaddad ™*O, vihang Wa 1, Reem Kateb? and Ahmed Bouridane #0 1 School of Ari Inligence, Guilin Univer of Electron Technology, Guin 541004, China The Laborstry for Imagery, Vision and Aric Intelligence, Ecole de Technologie Superieure, Montes, QC HIC 1K3, Canada College of Computer Scnce and Engineering, Tubsh Univers, Madina Centre fr Data Analytics and Cybersecurity, University of Sarah, Sharjah 27272, Ur Correspondence: [email protected] enor ahmadichaddad@gmailcom Abstract: The electroencephalography (EEG) signal is a noninvasive and complex signal that has ramerous applications in biomedical feds, inchiding sleep and the brain-computer interface. Given its complexity, researchers have proposed several advanced preprocessing and feature extraction methods to analyze EEG signals. In this study, we analyze a comprehensive review of numerous articles related to EEG signal processing. We searched the major scientific and engineering databases and summarized the results of our findings. Our survey encompassed the entire process of signal processing, from acquisition and pretreatment (denoising) to feature extraction, classification, and application. We presenta detailed discussion and comparison of various methods and techniques used for EEG signal processing. Additionally, we identify the current limitations ofthese techniques and analyze their future development trends, We conclude by offering some suggestions for future research in the field of EEG signal processing Keywords: FFG signal processing; machine learning OsKe Caton: Chia, A; Wy Ys Ka, 1. Introduction ctoecephlogaphy Sgn Brain science has become an essential field of study to unravel the mysteries of life due roca’ A Comprising to developments in biomedical technology and our growing understanding of the brain [1] Revuvand AnaipinofMtindsand Since the middle of the 20th century, the complexity of the brain has been studied and, in Teekiqus Snsrs203,25,6858 recent years, brain science has continued to be a hot topic for research [2], The analysis of Ips /Msonsomraniot rain electrical activity is one of the main ares of interest in bein science [3]. A esl, the cecroencephalogram (PEC) is important earls bran sence and is offen ed in various brainsclated research domalt [is] TEEG is a noninvasive neuroimaging techrique that involves the placement of cle Aes sy trodes on the salp to record electrical activity ofthe bran [e]. This enables researchers Revised: 16June 2003 to measure and analyze the electrical signals generated by the brain. These signals offer ‘Accepted 3 uty 203 valuable information on the operating mechanisms of the brain, covering the identifica- Pubs: 16 aly 200 tion of various neurological disorders and the exploration of cognitive processes such as perception, ltenton and memory EEG has gened widespread popular as means of Copyright 2 205 ty te suitors tistics [7] In addition, EEG signals have the potential tobe integrated with other imaging Conrhe © 2 5 Oe aor, lies, including agnete resonance imaging MRI [6] functional nea nfsred (Soe NOT Pe Seti troscopy (NIRS) [0=1>], and positron emssion tomography (PED 3, in order to TRS LESTE Sinewea beter understanding of brain funclion and stucture candle f the Creve Comment EEG is a signal patter that is obtained by amplifying and recording the spontaneous ‘tbton (CCBY) ken tips? iological potential ofthe brain on the scalp. This potential has been shown to reflect the ‘catecommons orp/bene/by/ macroscopic activity of the brain surface and is typically acquired using noninvasive cle a0. trodes applied onto the scalp. These electrodes capture the inherent and periodic electrical Sensors 223,23, 4, hepa / dot org/103380/s23146434 ‘pe /wrrwemépicom /joural/senaonsSensors 203,23, 6434 2ol27 impulses generated by clusters of brain cells [14]. Currently, EEG is widely used in the field of neuroscience and has the potential to advance brain-computer interfaces, facilitate emotion detection, and help in partial paralysis rchabilitation [15,16]. Furthermore, EEG is a valuable tool for clinicians and researchers in identifying brain dysfunction-associated diseases, including but not limited to Alzheimer's disease [17,18], epilepsy, schizophrenia, Creutzfeldt-Jakob disease [19], cerebral palsy (20], and cognitive impairment [21] ‘Accurately identifying and analyzing EEG signals requires a thorough understanding of their complex and theoretical properties, as well as the extraction of relevant features for the given task. However, EEG signals pose significant challenges due to their unique characteristics. According to [22], one such challenge is their susceptibility to noise interfer- ence, resulting in a low signal-to-noise ratio. Additionally, their nonlinearity and lack of conformity to a normal distribution distinguish them from conventional signals. Moreover, individual factors such as age, psychology, and testing environment can cause significant variations in EEG signals [23]. In [22], itis observed that the unique properties of EEG. signals pose a challenge in extracting pertinent information about specific tasks directly, from them. Therefore, itis imperative to develop various methodologies for signal anal and investigate machine learning techniques for signal analysis to better understand signals [24], Accurate extraction of zelevant information on specific tasks from EEG signals requires careful consideration of their distinctive characteristics and the advancement of, sophisticated signal analysis methodologies. As emphasized in [25], precise detection and analysis of EEG signals are crucial to advance our understanding of brain functioning, ‘The high interest in EEG as a research domain is apparent in the Google Scholar, PubMed, and Web of Science search results obtained between 2016 and 2022, as illustrated in Figure 1 (Our paper presents a novel contribution through a comprehensive description of denoising techniques, which includes mathematical formulations with pseudocodes. In addition, we report the recent advancements in the field of EEG, while highlighting current challenges and discussing future trends. The primary contributions of this paper can be summarized as follows. + Wepresent a detailed examination of the EEG signal analysis process, including the stages of signal acquisition, denoising, and feature engineering + The procedure used to denoise the EEG signal is described in full, along with the accompanying evaluation standards, + We examine feature engineering in detail in this paper, looking at time-frequency, high-order spectral, and nonlinear dynamic analysis ‘+ We givea thorough analysis of both traditional and deep learning methods for catego- rizing EEG signals. We also provide an overview of the typical datasets utilized for ERG signal processing + Wehighlight current issues with EEG signal processing techniques and offer potential solutions as well as future research prospects. The structure of this paper is as follows. Section 2 presents a brief description of the impact of EEG acquisition as a noninvasive biomedical device. It gives a full evaluation of various denoising techniques, and an explanation of the merits and downsides of each Additionally it examines the significance of feature engineering and its various techniques, as well as the use of machine leamning- and deep learning-based classifiers. In Section 3, future perspective and limitations are discussed. The article ends in Section 4, where the research efforts and contributions of this investigation are outlined,Sensors 2023, 23,6434 Sota7 7000 ‘Google Scholar |mPubMed = Web of Science # 400 E000 = 2000, hl hil al hil hil il il 2016 201 20820192020 oat 2022 Figure 1. Bar graph illustrating the number of published papers over the years obtained from Google Scholar, PubMed, and Web of Science searches for the subject of the lectroencephalogram (FEC) ‘The search query used was “EEG” inthe title 2. The Pipeline of EEG Signal Analysis In this section, the focus is on the methodology of EEG signal analysis, as illustrated in Figure 2. The pipeline is related to the classification of EEG signals. The discussion begins with an explanation of the methodology of FEG signal acquisition using equipment Then, the algorithm for denoising FEG signals is examined, which is capable of eliminating erroneous data and extracting relevant information. Next, the feature engineering process is discussed, which involves removing theless important features. Finally the use of deep learning and machine learning algorithms for classification is explored. Figure 2. The four-step process for EEG signal analysis, The EEG signal analysis involves four stages acquisition, denoising, feature engincering, and classification, 2.1, Acquisition EEG is a neurophysiological technique used to measure and quantify neural activity in various regions of the brain, The brain consists of a large number of neurons and their activities generate distinct scalp potentials, producing signals in different states of alertness, response to external stimuli, and other factors unique to each individual [26]. To obtain data sources for various applications and research on EEG signals, appropriate acquisition equipment is essential. Acquisition of EEG signals can be classified into two primary. categories: invasive and noninvasive [27]. Invasive acquisition involves surgical insertion of electrodes into the cerebral cortex or other regions of the brain to capture signals. On the other hand, noninvasive techniques use EEG sensors positioned on the scalp’s surface and do not require implanted electrodes. Currently, most EEG signal acquisition methods are noninvasive, In Algorithm 1, we provide the main steps of EEG signal analysis, During the initial stages of the acquisition of FEG signals, the German scientist Hans Berger detected clectrical signals inthe cerebral cortex via a galvanometer in 1924. After pre- liminary investigations, the scientists began implanting metallic electrodes directly into the cerebral cortex to capture comprehensive EEG readings. With advancements in computer technology, FFG signal collection techniques have improved significantly, leading fo a higher resolution of collected FEG signals. Most EEG signal acquisition instruments that have reached relative maturity currently employ PC displays, wired data transmission,Sensors 203,23, 6434 sot7 and extemal power sources. These devices exhibit robust data processing capabilities, favorable outcomes, and consistent performance. However, they have a heavy form factor, pose potential hazards, and exhibit high power consumption. Consequently, the need for portable EEG collection equipment has been observed [23] ‘Algorithms I Pipeline of FEG signal analysis Input: EEG signal eeg(0). procedure Acquisition (eeg(): Sampling equipment selection. Choose invasive or noninvasive acquisition. Choose the best reference electrode, Return: EEG data eeg(() procedure Denoising (eeg()): Using different denoising methods to process EEG signals, Return: Denoised EEG signal eeg (tn procedure Feature Engineering (eeg(t) Perform time-frequency, high-order spectrum analysis or nonlinear analysis. Return: More expressive features, eeg() procedure Classification (eeg('), choice) Perform classification tasks based on choice: If choice equal traditional approach do Using KNN, SVM, .. to classify, If choose the same deep learning approach do Using CNN, GAN, .. to dlassify Return: Accuracy, precision, ete Portable EEG acquisition devices such as the Emotiv EPOC have gained popularity in recent years, The Emotiv EPOC utilizes nonimplantable electrodes and comprises 14 electi- cal dats acquisition channels and two reference electrodes, Stytsenko etal. [29] found that the Fmotiv EPOC can collect real FEG data, Emotiv FPOC neural headphones are also avail- able on the market to measure brain activity [30]. However, the performance of the Emotiv EPOC headset devices, while capable of recording EEG data, is inferior to that of larger devices [81]. In another study, Martins proposed a wearable EEG acquisition device and a sleep inertia detection system of the data analysis platform [32]. The system is a monolithic low-power with a low-noise analog frontend EEG acquisition system [33]. The system has demonstrated high precision and high reliability, and flexible adjustment. Moreover, a new ‘waterproof, lightweight, and portable EEG acquisition device was proposed to acquire and analyze the EEG signals of dolphins [34]. The device was designed to enable relatively unrestricted FEG acquisition. Their acquisition device is equipped with customized suction cups with embedded electrodes. It also incorporates a Bluetooth module to communicate with the ground station. Furthermore, they used the portable Muse brain wave sensor device for stroke identification [35]. The device follows the intemnational 10-20 system and utilizes four recording electrodes (AF7, AF8, TP9, TP10) and one reference electrode pz). In [36], they designed a high precision portable EEG acquisition system using the Compact?CT platform to address the limitations of existing EEG acquisition systems, such as high costs and limited accuracy. Aside from the previous methods, there are many ways to collect EEG signals. For example, deep brain stimulation was performed through the ¢ of neural electrodes that were placed in specific target regions ofthe brain [37]. These electrodes generate current or voltage through an implantable pulse generator. Further- more, the MR signal has the ability to reflect both oxygen saturation and blood flow in the brain [38] It can reflect the activity of neurons and serve the purpose of functional imaging,Sensors 2023, 23,6434 solar Figure 3 shows a short segment of an BEG signal [38] recorded using a ld-channel Emotiv Epoc device. These 14 EEG signals are specific locations on the scalp where electrodes are placed to measure brainwave activity. In general, the frequency of an EEG signal is measured in Hz and refers to the number of cycles that occur per unit time [10] Specifically, there are five waves that correspond to five ranges of frequencies in EEG. signals [41] namely, Delta [0.5 to 4 Fiz], Theta [4-8 Hz], Alpha [8-13 112], Beta [13-30 Hz], and Gamma [>30 Hz]. Delta waves are related to slow-wave activity in the brain and deep sleep; Theta waves are commonly observed during periods of relaxation and meditation; Alpha waves are most noticeable when the eyes are closed; Beta waves are associated with cognitive processing that is actively engaged and can be detected during tasks that require significant attention; and Gamma waves are linked to advanced cognitive processing and the merging of sensory information. In addition, each EEG channel is named according to its location relative tothe midline ofthe head and its distance from the forchead or back of the head. Together, these channels allow researchers and clinicians to measure electrical activity in various regions of the brain, and gain insight into cognitive processes such as attention, memory, and emotion. | | : Time Figure 5. An example of 14 KEG signal channels, where the xaxis denotes time and the y-axis represents the magnitude of the 14 signals [39]. These signals can be characterized by their frequency, ‘which refers to the number of eycles per second (Hz) ofthe electrical activity. These channels are named AF3,I7, F3,FC3,T7, P7, O1, 02, P8, TS, FCS, F4, FB, and AFA, which correspond to the specific clectrode placements on the Emativ Epoc equipment. ‘Table 1 presents a summary of recent studies on EEG signal analysis in different application domains. In [42], commercial EEG equipment was used to collect signals to analyze driver fatigue. A third-order Butterworth bandpass filter was used to remove irrelevant information from the raw signals [42]. Another study by Gamage etal. utilized the Emotiv EPOC to detect the driver's emotions while driving and extracted features using EEGlab and other tools in Matlab. The data was then classified to develop an emotion classification model aimed at reducing the risk of car accidents, The study demonstrated the high reliability of the dataset collected by the device [4]. Table 1 provides relevant details, such as the study objective, data source, and data processing techniques used in the different studies,Sensors 223, 23,684 sot20 ‘Table 1. Research on EEG applications. Re Prapors ‘Acquistion Method Dut Processing Ma (2) Recognize drive fatigue Gammel Neon om ih Rr Batak Gamage (61 Detect divers EECHo reduce tafe Brake the cmations ofthe test diet cL AB Tplbox of Mat Seiten the depression Tndilonal 12¢lecteode mounted hen [4 8 pr clastic cap and s wearable electrode EEGLAB Toolbox of Matlab ‘eeogiton performance elastic cp ad Saedi [45] Defect the working satus of Investigate mental and motor work A mix of macro and micro scrutiny constriction workers Han [16] Classification. state [EEG measured around the ear stimating classification accuracy ts} ‘ ot eye “ whe ‘using 3 CNN models Paw and Parke [17] Use BCI contol the robot Emote EPOC a Chen (01 ERG decoding Obtained inthe open world Siperised deep ening Pa [sl Pre electrode in BCL Obtained form PcG lcd SSUE based bet Seal pleptc etre prediction Publcy availble CHBMM dataset Deep neat network mode Wen [51] Evaluate spatial cognitive ability _‘FYOm7 subjects participatinginthe — Csuoting strength calculation Li s2) Emotion recognition Expetimentlevel BN Frlaiath ta 59] ‘Trestnent and diagnos of ADHD HILCES framework ‘CNN: Convolutional neural network: BCE Brale-machine ntact: BN. Batch normalization; ADHD: Attention debt hyperactivity disorder, HILCPS: Humaninsheloop eyberphysiel systems 2.2. Denoising [As mentioned above regarding the acquisition of EEG signals, multiple electrodes are placed on the scalp. However, external interference can cause diverse artifacts to emerge, which can compromise the quality of the signals. Physiological artifacts, such as involuntary eye movements, blinking, heart activity, and muscle movement, are known fo be present in FEG signals and can negatively affect their quality [54]. Therefore, denoising EEG signals has become a topic of significant research interest and attention. To ensure the reliability of features extracted from EEG signals, itis essential to remove any associated artifacts, Currently, several denoising techniques have been developed, 2.2.1. Regression Method The traditional approach to remove eye artifacts in EEG signals isthe regression- based analysis approach. Duting the recording of the EEG signal, an electro-oeulogram is recorded concurrently to obtain coefficients for various noise sources such as blinking artifacts (VEOG), eye movement artifacts (HEOG), and other noise sources, by using regression analysis, These coefficients estimate the ratio of artifacts in a particular EEG channel [55]. In [56], they proposed a general lagged regression model to represent this process a follows. ceg(!) ~ EEG() ~ Lo Be coglt —8) o where eeg(t) and eog(t-g) are the recorded EEG and E0G information at times ¢ and fs, respectively. EEG(t) denotes the uncorrupted EEG at time f, The regression coefficient By measures the effect of EOG on eeg() at time (1g) For example, to remove the effect of an artifact signal Y from an EEG signal X, we compute the regression correlation between the two signals. This denoising function allows tus to successfully eliminate the artifact signal from the EEG, Algorithm 2 illustrates the general procedure. However, the main concern raised about the regression method is bidirectional con- tamination [55]. For example, FOG recordings may include neural potential along with ‘ocular potential [57,58]. Therefore, removing EOG activity from the EEG signal requires subtracting a portion of the relevant EEG signal from each recording. Furthermore, one of the challenges of regression techniques is that they may not be effective in dealing with other artifacts, such as EMG artifacts [59], due to the lack of lear reference channels. With the emergence of potentially more efficient algorithms such as principal componentSensors 208,23, 6434 Pol27 analysis (PCA) and independent component analysis (ICA) [60,61], the regression method is no longer the default choice for removing artifacts from an EEG caused by EOG or ECG signals, ‘Algorithm 2 Regression-based denoising of EEG signals Input: EEG signal X, artifact signal Y Output: Clean EEG signal Z function REGRESSION(X, Y) Calculate regression coefficients between X and Y Remove artifact from EEG signal return Clean EEG signal Z end function 2.2.2. Blind Source Separation Blind Source Separation (BSS) i a technique that separates source signals from a mixed signal without prior knovledge of the original signals. Inthe first step, the observed EEG is decomposed int its sources using the BSS algorithm, The noise sources are then identified and eliminated while preserving the sources of brain activity. BSS techniques are commonly used to denoise EEG signals [6] The following techniques are described below: Independent Component Analysis: Various types of ICA algorithms have been pro- posed in different papers, making it one of the key blind source separation techniques in biomedical engineering, ICA is capable of extracting statistically independent sources from. a collection of recorded signal [63]. In general, denoising EEG signals using ICA can be expressed as follows [64]: X= As @ where 1e given equation involves three matrices: X, A, and S. Matrix X contains EEG data, matrix A represents the linear mixing of various sources (e.g., EEG and artifact sources), and matrix S consists of independent components, such as brain and artifact sources. After obtaining the independent components, they can be visually examined to identify any artifacts such as eye blinks and muscle activity. The EEG signals can be denoised by removing these components. Algorithm 3 presents the denoising EEG signals based on the ICA method, ‘Algorithm SICA based denoising of EEG signals Inputs X: EEG data matrix Input: 1_components: number of independent components to estimate Output: 5: matrix of independent components Output: A: estimated demixing matrix Center and whiten the X Initialize A randomly. repeat Update A by exploiting non-Gaussianity of independent sources until convergence Compute $ from A and X. Identify artifact components in S. Remove artifact components from S. Reconstruct cleaned data from S. return $, A Due to the significant overlap between EEG signals and EMG artifacts in both spatial and temporal domains, conventional ICA algorithms often struggle to separate all EMG artifacts and generate a set of independent components. To address this problem, Li etal introduced an improved ICA model called EMG removal by adding sources of EMG (ERASE) [65]. Specifically, EMG reference artifacts were involved from the head and neckSensors 203,23, 6434 sot27 muscles as input to the ICA, which increased the power of the EMG artifacts in a few independent components, resulting in a more accurate separation, In comparative tests, the ERASE algorithm removed an average of 26% more EMG artifacts from EEG signals than the traditional ICA algorithm. In [55], a new denoising framework was devised and merges ICA with the continuous wavelet transform (CWT). CWT and the K-means algorithm are used to map the detected blink artifact. EEG information is then preserved swhile denoising through ICA [65] Principal Component Analysis: PCA, which stands for principal component analysis, is an easy-to-use data reduction technique that uses the principle of orthogonality to climinate artifacts [54]. Through the utilization of PCA dimensionality reduction, it is possible to eliminate the presence of noise, represented by small eigenvalues, within the data, This process results ina partial denoising effect. Typically, given EEG data X, the main goal of PCA is to solve this equation: XXT ey = Ait 8 where the symbol A denotes the eigenvalue, while w represents the eigenvector. The tech nique for reducing the number of dimensions in a dataset is accomplished by decomposing the eigenvalues of the matrix XX". The resulting eigenvalues are then sorted and the top d values are selected to serve as a projection matrix. Subsequently, the projection matrix can be used to transform the EEG data denoted by D into new EEG data represented by D" = W"7D, while minimizing the presence of noise, The pseudocode for PCA is presented in Algorithm ‘Algorithm 4 Typical principal component analysis Input: EEG data D ~ {1,23 4y} low-dimensional space dimension Output Projection matrix W! (ta) procedure PCA(D): Sample centering x) x) — 2 O% # Calculate XX7. Eigenvalue decomposition for XX" Select the largest d eigenvalues WE (ay 22,124) New EEG data D* = W*"D. Retum: D* In recent times, thet has been an increasing focus on utilizing principal component analysis in conjunction with other techniques to achieve EEG denoising. Patel et al. have demonstrated the effectiveness of combining ensemble empirical mode decomposition (EEMD) with PCA to efficiently detect and suppress artifacts in single-channel EEG data ‘This method can automatically detect and suppress c the detection threshold [67] In [63], a learning model based on PCA and semi-supervised support vector machine (SVM) is proposed. The model first preprocesses the EEG and uses PCA to reduce its dimensionality: After obtaining 2 set of optimal channel subsets, a semi-supervised classification model based on SVM is designed. This model determines, the relationship between labeled data and unlabeled data by calculating the Euclidean, distance between them, and then extracts features to identify them. The experimental results indicate that the method can achieve 84.3% corzectclasifiation results with only 40% labeled data, suggesting its potential in scenarios where only a small amount of labeled. data is available [63] e artifacts after correct selection of, 2.2.3, Canonical Correlation Analysis In the context of EEG signals contaminated with muscle artifacts, eanonical correlation analysis (CCA) is generally more effective than ICA [59]. Due to the relatively lowerSensors 208,23, 6434 90t27 autocorrelation of muscular artifacts compared to brain activity, itis feasible to employ canonical coreelation analysis (CCA) asa means of distinguishing between musele activity and brain activity [70]. The common CCA formula used for denoising EEG signal is expressed as follows [71] sma Ras? Fata Rye where the variables and v represent canonical variates, which ate linear combinations of channels. The EEG data and artifact data have covariance matrices denoted by Rex and Ryy, respectively. The cross-covariance matrix between the FEG data and artifact data is represented by Ryy. After obtaining the canonical variates, they can be used to climinate artifacts from the EEG data by subtraction. The denoising steps based on CCA are presented in Algorithm 5. a ‘Algorithm 5 CCA based denoising of EEG signal [72] Inputs X: EEG data matrix Input: Y: matrix of auxiliary variables (e.g,, EOG or ECG data) Output: Z: matrix of cleaned EEG data Center and whiten the X. Initialize the weight matrices A and B randomly. repeat Compute the canonical weights w, by maximizing the correlation between X and Y with respect to A. aoe: Weights used to linearly combine the EEG signal for one component ‘Compute the canonical weights: by maximizing the corzelation between X and Y with respect to B. > a:weights used to linearly combine the ausiliary signal for one component Update the weight matrices A and B. runt convergence Compute the cleaned data as 7 = ATX. retumn Z In [69], P Sheoran et al, proposed a new algorithm that combined CCA and noise adjusted principal component transform (NAPCT) to eliminate noise in EEG data. Using CCA to estimate the noise covariance matrix and NAPCT to remove artifact components, the algorithm achieved this without human intervention [69]. Another study introduced an unsupervised automated eye artifact recognition and removal algorithm [73]. This algorithm used CCA to extract neural signals from data and used a multi-channel Wiener filter (MWF) to adaptively climinate eye artifacts from multi-channel EEG data [73] 2.24, Wavelet Transform The signals ecorded by EEG devices often contain iregularities. To analyze these non- stationary signals, the wavelet transform (WT) isa widely used method [74,75]. The con- ventional approach of WT divides the EEG signal into wavelet components. Components that contain artifacts are identified and removed, leaving only clean components. These clean components are then used to reconstruct a purified signal [76 The WT is generally classified into two categories: discrete wavelet transform (DWT) and continuous wavelet transform (CWT) [77] Given the continuous nature of the FEG signal, our attention is directed towards the DWT transformation. The DWT can be expressed in mathematical notation: DWT(mn) = [1S x(t) Pmn(tat ® ‘where the variables "m” and “n" represent the scaling and translation factors, respectivelySensors 203,23, 6434 awol27 The process of discrete wavelet analysis involves the decomposition of x(t) into various scales: LL {yal + Le axieca(t) © where the given equation involves discrete analysis wavelets represented by s(t) and discrete scaling functions represented by ¢x4(#), The variable d;(X) denotes the detailed signals or wavelet coefficients at a scale of 2), while ag (k) represents the approximated signal or scaling coefficients at a scale of 2. DWT is a method of transforming time domain EEG signals without redundancy, ‘which is useful in removing artifacts. To accomplish this, the signal undergoes a series of low-pass and high-pass filters to obtain approximate and detailed coefficients. This process is repeated until the desired frequency is achieved. In Algorithm 6, we present the pseudocode for DWT. However, DWTT has the drawback of lacking translation invari- ance, The stationary wavelet transform (SWT) can overcome this issue, but it has its own limitations, such as redundancy and slow speed [78] ‘Algorithm 6 DWT based denoising of BEG signal [9] Input: X: EEG data matrix (rows represent the EEG channels) Output: Y: matrix of cleaned data Set the wavelet basis and level of decomposition for each channel cin X do Compute the DWT coefficients of ¢ at each level using the fatigue wavelet basis. Identify the approximation coefficients at the desired level as the artifactfree signal ‘Threshold the detail coefficients using a soft or hard thresholding technique. Reconstruct the cleaned signal by inverse DWT using the modified coefficients. Store the cleaned signal in the corresponding row of Y. end for return Y WT alone may not be suficient to address all the issues associated with FEG signal denoising, as it can resultin information loss and signal reconstruction problems. Therefore, the combination ofthe wavelet transform with other techniques to improve the denoising ‘process has been explored. For instance, the authors of [80] applied ICA to separate signals based on WT and found that this combination was effective in removing EMG noise and ECG artifacts from FEG signals In addition, notch filters can be used in conjunction with WT to address the issue of overlapping spoctra/ frequencies between EEG signals and artifacts. In [51], an adaptive threshold for wavelet coefficients was used to eliminate frequent ocular artifacts (OA) and a 50 Hz IIR notch filter to reduce artifacts and noise while preserving the original brain signals 2.2.5, Empirical Mode Decomposition Empirical mode decomposition (EMD) is a technique for analyzing non-stationary and nonlinear signals that offers several desirable properties. EMD leverages the signal extreme points to decompose a signal into a set of intrinsic mode functions (IMEs) and a ‘monotonic residual, which can be expressed using the following formula [82}: x) = PGW) +i) a where C, represents an IMF and ry represents the monotonic residual. The IMFs are capable of capturing the fundamental oscillatory components at various frequencies, which facilitates the differentiation between artifacts and the intended EEG signal. To obtain the high-frequency content of a signal, including any artifacts, as a detail component, one can subtract the envelope from the input signal. The envelope representsSensors 203,23, 6434 nota the smooth curve that passes through the local maxima and minima of the signal. This technique enables the elimination of undesired artifacts while retaining the original EEG signal. The reconstructed signal is obtained by adding up the detailed components after cleaning, More details are listed in Algorithm 7, ‘Algorithm 7 Empirical Mode Decomposition (EMD) for EEG Anifact Removal [82] Inputs X: EEG data matrix Output: Y: matrix of cleaned data Set the stopping criterion and number of maximum iterations. foreach channel cin X do Initialize do = c, k= 1 ds signal at iteration & repeat Find the local maxima and minima of dy. Compute the envelope by interpolating the maxima and minima, Subtract the envelope from dy_; to obtain the detail component hy. © I: detail component Update dy = dy ~ by. Increment k ‘until stopping criterion or maximum iterations are reached. Compute the reconstructed signa src= Sh yh >: reconstructed signal for channale Store the leaned signal in the coresponding row of Y end for return One of the advantages of EMD is its ability to extract local amplitude, phase, and fre quency content from the resulting components. EMD is also adaptive and efficient and, when combined with other techniques, itcan lead to new advancements in the denoising. of FEG signals. For instance, combining ensemble empirical mode decomposition (FEMD) swith the CCA technique led to feasible results. Specifically, EMD generates a large num- ber of IME, increasing the number of channels available for ICA. This method leverages interchannel information and addresses the challenging problem of CCA in dealing with FEG data with low signal-to-noise ratio (SNR) and complex contamination [83]. EMD is highly sensitive to spike noise because of its reliance on extreme signal point features for IMF decomposition. This sensitivity can lead to the mode-splitting effect, which can seriously affect the removal of EOG artifacts, To address this issue, the multivariate adaptive moving average-empirical mode decomposition (MAMA-EMD) based method extracts peaks into the first IMF to improve the accuracy of sulbsequent IMF screening and alleviate the mode-splitting effect [84]. However, MAMA-EMD may not achieve optimal results in separating spikes when the pulse has two or more consecutive spike points, ‘fo address this limitation, a new version of MAMA-EMD is proposed by supplementing the minimum arc length criterion, This approach effectively eliminates the influence of, multi-point spikes on the screening process [85]. In recent years, various classic and commonly used denoising methods have been combined to achieve better signal denoising in different situations, Some examples of these methods or their combinations that have been used for denoising puxposes are summarized in Table 2. 2.3. Evaluation Criteria for Denoising Various metrics, such as mean squared error (MSE), root mean squared error (RMSE) signal-to-noise ratio (SNR), and percentage root mean square difference (PRD), are com monly used to assess the effectiveness of EEG signal denoising [86]. MSE is frequently employed to evaluate the similarity between the initial EEG signal and the noise-reduced signal. The RMSE is mathernatically defined as the square root of the MSE. The SNR is a metric used to compare the magnitudes of the signal and noise power. The PRD is used toSensors 203,23, 6434 wet measure the degree of similarity between the original and noise-reduced signals, with a lower PRD indicating a higher degree of similarity between the two signals. These can be defined as: rics @) where y, typically refers to the true EEG signal value at time point i, 9, represents the denoised value of the EEG signal at time point i, and mis the total number of values. Table 2. Application of denoising method in EEG signal. ca Signa ecg Metiod Condon wie EG heenee ares of nck and had muy MET eG EMC pation wou maa Madina nd Wnvla [6] CWTand Keane ee ae Rutomatederton Sn suppres of aman eye as Combining EEMD and PCA flocs cm bese sont PREP OED Xie [68] ea aera ear signal-to-noise ratio and only few labels, with high Sheoran and Sain [59] Mino 173) Zao and Gotmnan [0 Tibdewal (61) chen (83 Yang {641 Land Zang (5) classification model Secogniton sccursey and les traning time Artifact cormponenls are removed withost ‘Combining CCA and NAPCT Acta com Eye artifacts canbe adaptively removed ftom CCA and MWE _mult-channel EEG data without the need fora neference signal ‘The combination of wavelet transform and ICA can Hfcively remove EMG snd ECG atface in EEG signals Execively reduces artifacts and nous wile preserving, the original brain signal. Teen make good use of iterchannel information and nas good afact removal effect inthe eee of eros signal polation Cap allevste sting effects, but not suitable for separating maltipoin opis Itean eliminate he effect of maltpoint spikes on IMI screening and better remove BOG arcs Wavelet anaform {Use the adaptive threshold of wavelet coeficents EMD and CCA techniques Extract pikes othe ist ME =MD TCAs Indpendent componantsnlyain PCA: Prnalpal component analy CCA: Canonical conlaon saya WE Wavelet transform EMD: Empirical mode decomposition, CWT. Continuous wavelet transform, EEMD: Ensemble empirical mode decomposition; NAPCT Noise adjusted principal component transform; MWE: Malt 24, Feature Engineering Enhancing feature engineering can improve the accuracy of predictions made on raw data by transforming them into more expressive features, Extracting the availble features from the processing of EEG signals is a complex task that typically requires multiple human, experts with specialized knowledge. Machine learning techniques, such as deep neural networks and adversarial generative networks, have allowed the automated extraction of features from EEG signals. However, the interpretability problem of deep learning is deeply criticized. Recently the advancement of explainable AI (XAI) methods [87] has aimed at improving the interpretability of deep learning models. For example, XAI based Smooth- Grad [8s] is used to perform EEG based emotion recognition [8], seizure detection [9], and other applications. Using XAT methods, it allows eliminating the need for manual tion by human experts. In [I], the authors proposed a method to achieve high- resolution assessment of neural activity using deep networks, involving the implementation of relevant layer propagation. They used an adversarial generation network to produce EG signals [92], However, efficient feature engineering methodologies can help machineSensors 208,23, 6434 w0l27 earning models acquire the fundamental features implicit in EEG more easily. Therefore, we examine the main feature engineering methodologies used in EEG [91-93]. Commonly utilized conventional signal processing techniques in various research studies on EEG signal processing include time-frequency analysis, high-order spectrum analysis, and nonlinear dynamics analysis. 24.1, Time-Frequency Analysis ‘The primary purpose of time-frequency analysis is to establish a link between the time and frequency domains. This involves analyzing and processing signals in both domains to extract relevant features, The most commonly used methods for analyzing stationary signals include analysis of variance, waveform parameter analysis, wave identification, histogram analysis, correlation analysis, and others. These methods are often applied in the diagnosis of diseases (94. For example, time-frequency analysis is used to map EEG signals in the time and frequency domains. By dividing the signal data into windows and scoring them, itis possible to identify epilepsy signals in the time-frequency domain [94], ‘Another example is related to the detection of peak features in EEG signals [95]. Specially, the signal is denoised using SVD and then the peak is detected, which yields superior results [95]. In [96], a high resolution non-parametric time-frequency method is proposed to analyze EEG signals that uses CNN to optimize the Wigne:-Ville distribution as an input ‘without parameters, showing its clear superiority over other methods, 242. High Order Spectral Analysis Although the time domain analysis method falls shortin analyzing high-order informa tion and providing complete signal feature results, the high-order spectral analysis method can effectively address this limitation [97,95]. With its ability to map specific information rote effectively in EEG signal processing, higher-order spectral analysis demonstrates significant superiority [97]. Itis capable of suppressing Gaussian noise and producing spectral structures that reflect more information, as evidenced by simulation experiments con more than 200 EEG samples conducted by some authors [97]. The resulting spectral lines are Matter, with ess noise and smoother contours (27]. To identify the nonlinearity and high dimensionality present in epileptic signals, the principal component features are extracted using PCA on the 15 high-order spectra (HOS) features extracted from the FEG data [99]. Furthermore, in [100], they use HOS to analyze FEG signals in the field of neuro-marketing. According to the study findings, the proposed model, using SVM. with Gaussian kemel, achieved an average accuracy of 73.24% across all users. Moreover, HOS features were used to access participants who were in atypical emotional state but not exhibiting any motor movernents [10]. Experiments demonstrated that the method achieved an average accuracy of 95.7%. 243, Nonlinear Dynamic Analysis Previous work has shown that the traditional linear analysis method is inadequate for accurately evaluating the dynamic structure of EEG signals and as a result it cannot reveal the essential characteristics of brain activity: However, the use of nonlinear dynamic methods to extract and analyze EEG signals has provided a new approach to further study the process and characteristics of human brain activity (102] Much literature inthis field uses nonlinear techniques, such asthe Lyapunov expo- nent, complexity measures, and fractal dimension, to analyze EEG signals. For example, a researcher used the correlation dimension of nonlinear dynamics and the Lyapunov index to extract characteristics of high-frequency EEG from elderly and young subjects during various activites, including silent eye closure, silent eye-opening, and N-back letter memory events. They then performed a statistical analysis on the resulting eigenvalues to compare the differences between the two groups of eigenvalues [103]. “The features ofthe EEG signals were effectively extracted by various indicators such as the Hurst index, the Lyapunov index, the sample entropy, and the wavelet entropy [103]Sensors 208,23, 6434 aWot7 Furthermore, an adaptive Lempel-Ziv complexity algorithm was proposed and utilized to measure the complexity of EEG signals, which was capable of identifying emotions [105] By comparing the complexity values ofthe traditional Lempel~Ziv-Welch compression algorithm (LZC) and the multiscale and adaptive LZC algorithm on the corresponding electrodes under three emotional states, it was discovered that the adaptive LZC algorithm could effectively distinguish between the three different emotional states. The algorithmic processes of permutation entropy and sample entropy were briefly introduced and their respective advantages and disadvantages were analyzed in detail [106]. A new algorithm called equal symbolic entropy (ESE) was proposed, and its effectiveness in terms of accuracy and efficiency was verified through simulation. It was applied to analyze emotional human EEG signals from an experimental group [07] Furthermore, due tothe intricate nature of the EEG signal, traditional linear techniques face significant challenges when analyzing it, leading to the application of nonlinear dynamics methods in the analysis of EEG signals. Various features such as the correlation dimension, the fractal dimension, the complexity, the approximate entropy, and others have been explored in the literature, along with the corresponding methods for extracting these features from EEG signals. These findings provide evidence supporting the scientific basis of utilizing nonlinear dynamics for EEG signal analysis [108] 25. EEG Based Classifications EG signal classification is a fundamental task in the analysis of brain function, which can be considered as one-dimensional biomedical signal processing [10%11]. Various processing methods can be employed to classify EEG signals, including statistics, machine learning (deep learning), and other techniques. In this section, we will focus on the classification methods used in various fields of EEG application, with particular emphasis ‘on machine learning-based approaches. Furthermore, we have listed the commonly used EEG datasets in Table 3, Table 3. Summary of commonly used EEG signal analysis datasets, Data Sample) Type ‘Sram Zhang (113) 1 Dhjec recognition oust (113) 2 Emotion analysis rang and bu (1) ig Emotion recogrition ‘Ang (115) s Emotion eecorition Tangermann [161 4 cl Saja 1071 5 Bel Andragak [118] ° Setrre detection ‘Shosb [15] B ‘Seizure detection/ prediction Dest 12 4 Suzure detection precicion Zrang (121) 6 Mental werKosd Venkatachalam [122] 5 Mic Zrseg [23] 6 EEG denoising 5 [BET Brin computer ntrtces MIT Motor imagarycnsaicationy SF Sample Reueney 25.1, Traditional Classification Method The classification of brain signals through the application of ML techniques mainly involves the use of supervised and unsupervised learning methods, These methods include naive Bayesian (NB), decision tree (DT), K-nearest neighbor (KNN), support vector machine (SVM), and random forest (RF), among others. Supervised learning is related to using input and anticipated output data to develop predictive models that cater to classification and re gression, On the other hand, unsupervised learning involves proposing a prediction model that uses input data for clustering and dimension reduction [124]. Supervised learning yields higher accuracy than unsupervised learning when using classifiers such as SVM or KNN. The precision of a solitary classification technique is restricted to particular use cases, Hence, multimodal integration algorithms are commonly employed in studies to improve the overall accuracy of classification. The increasing complexity of algorithms may result in bias and affect their accuracy. Previous research has used machine learning methods toSensors 208,23, 6434 15 0f27 examine FEG signals for the identification of diseases (e g., epilepsy, depression, stroke) and rehabilitation interventions (e.g., motion imagination). Specifically, Table 4 summa izes literature examples related to epilepsy, motion imagination, depression, and stroke. In general, the results in Table 4, suggest that the SVM model is an important classifier model for EEG signals. ‘Table 4. Application of traditional classification method in EEG research Re Domain Propose Method Conclaron andA To Tpikpey Us# KNN and ANN cases to predict For KNN dassfsy HAD with ample etopy had te * 1 PUP ious Iighest accuracy of bout 96% Pag 141 ptepsy ised SVM csr based noninetfSygsttlymproed te coet copnin te Jamunadei (17) Fpipay Used Ror dteston and esti IF had beter ress inclinnting eplepey eer deter 2a ‘Added. Gaussian noise and performed binary The maximm average cassicaton accarty of KN LeOUED Ml Gsesfeston clase ached 88507 ving 51 ar __Creled slower extent Mlclasidcaton algo The average dasiiatonscurcy of th to paradigms 8 ‘thn based on LDACKNN. was 6750 and 642% rspecvely Dengare and Padole Chet a oenty ving BBA NA GRBAN a GCRS re MI SVMULDA'and ANN aes Cen Ren (1) Stroke Adopted C43 decison 56 Consiacted « DT model with 37 nodes Haniwenand Yona] Stoke Used HOC and AUC formodelserering _The SVMmodel performed bes as AUC= 1000 tone feed State Build mode! Dased on LASSO, DWI PAL and The acuracy ofthe combined model was 0 822, which svM swasbetter than that the single sequence model “The andom fret nodal ad the bent prediction effect, Yong 134) Stoke Used CTA image collection et data and fold 8 sary of 99% aid 908% ta predicting new caoss validation Jachemie stroke DWE: Disest wavelet transforms SYNE Support vector machines; KN K-ncares acighbor; ANN Aili neural network, HED: Hight fractal dimenor: RI Random forest ME Motor tmagery: LDA: Linea deci imant analysis FIR: Finite impulse response; SSA: Sparrow search algorithm, KS. Kelmogorov-Smimow, NB: Naive Bayesian, MEDIA’ Murtiratl declined fuctation analysts DT Decision tree; ROG Receiver operating characteristic AUC: Area under ROC curve, DWE Dison weighted imaging: PWL Perfusion weighted na CTA: Computed tomographic arteriogsaphy. We also summarize many publications in Table 5, related to FEG signals to evaluate many other tasks. For example, ANN and KNN classifiers are used to classify the EEG signals into epileptic and non-epileptic classes. The results showed that the combination of statistical parameters and classifiers achieved an accuracy of 95.9% for ANN and 92.4% for KNN, indicating the effectiveness of the proposed method in detecting epilepsy using EEG signals [125]. FEG with an SVM classifier based on nonlinear feature extraction is used to improve the recognition rate of epileptic brain signals, They decomposed the EEG signal into different frequency bands through a four-layer WT and the approximate entropy (ApEn) value of the wavelet coefficients in each frequency band was used as the eigenvector input. The correct recognition rates for normal and epileptic FEG signals were 98.3% and 95.6%, respectively, which outperformed other similar algorithms [126] Furthermore, the results showed that RF performed better than AdaBoost and KNN in liminating ertor detection [127] Furthermore, there is extensive research on classifying EEG signals in motor imagery (MD, a process that involves imagining actions without physically executing them. This research is important in helping patients who have lost motor function in recovery. In {128}, it proposed to add Gaussian noise to EEG signals to improve recognition rates and perform binary classification (left hand and right hand) MI tasks. The KNN algorithm achieved a maximum average classification accuracy of 88.57% [129]; the LDA algorithm was used to reduce the dimension of the feature data in the classification of MI EEG signals. When combined with KNN, they achieved average classification accuracies of 67.5% and 84.62%, respectively. This improved classification accuracy and speed, demonstrating the algo- rithm’s advantages in lower-limb MI classification. In [130], the mixed feature majority vote classifier is used to recognize MI EEG signals. They selected the combined feature as the input of the classifier and used most voting classifiers to combine SVM, LDA,Sensors 223, rwol27 and ANN. The accuracy of performance measurement reached 85.36%, indicating that the proposed system outperformed the conventional machine learning EEG recognition classifier In [135], they expanded the use of BCI to include motor imagery and presented a framework that used augmented covariance extracted from an autoregressive model for classification purposes. ‘Table 5. Summary of various EEG stuclies ina variety of applications. Re Domain Proposed Method Conctasion Alka and Alot [185 Parmar [157] Ling snd As [138] Hossain [135] Padayaty spd K [260] YYoohu and fraiang [43] Shui [122] Satyanaeayans [265] ‘Proposed Hamming window bandpass F eo ‘ker model for automate gender denestion using classifiers Evaltated the performance of thé nonlinesr The maimnath accuricy rate of RBF kernel The RF clssiier based on negative emotion [BEG signal had the lowest ero peerage Dye KStefSva Conran nd ROE, vv had en th ptcae bea eee pepe vet EefurmedamudecnweSscher het dain acungrech Combing rangi oe The ecco KN ws 4% ate ‘sed new method decode Engh i a ites dc om BE gle Sindhi tino SRipeuted eles petamance wh sean Desnofasuiable chasers dsngush tone csteston nif as S2 ‘Schizophrenia ‘SZ EEG signals from HC EEG signals and an overall accuracy rate of 89.58% for enn tha coded erigevaiy Custom verigo sates on The RF nde tbat cient tiga state ‘machine learning and BEG signal analysis with an accuracy rats of &2 5% Alcoholic ‘Used NTFT and Kross validation method NN elsif achieved good rsultsin eee wh ee maione ‘An Svidon duierbscdon a Timescale cen detecting emotions “GD. Gander dctections FR Finite impulse response RSF Radial asi anton, BC: Brain-comptar interface Schizophrenia; NIFT: Normal time-frequency transformation; Hc: Healthy contol, CR Character recognition; IBF Radial basis function, ICH. Ineacerebral homahage, HRV: Heart rate variability; RESP Respiratory 16, DRA. Dedined Buctusuon analysis (ce other abbrevialions noted in Table ) The advancement in EEG classification models has led to new possibilities in detecting depression, a mental disorder that affects a significant portion of the global population. To recognize depression based on EEG signals, some researchers have utilized the tree rmodel's feature selection algorithm to establish a depression recognition model. Itis clear from Table 5 that research on the use of traditional ML algorithms to clasify EEG signals is still ongoing and is expanding. EEG signals are complex patterns of electrical activity in the brain, and accurately classifying them can be crucial to understanding various neurological conditions and cognitive processes. everal research studies have utilized entropy measurement and statistical features of EEG signals in gender detection to enhance its accuracy. To obtain EEG data on negative and positive emotions for training and testing, a finite impulse response (FIR) filter model is commonly employed. Decision trees, random forests, and multi-layer perceptron are popularly used to predict gender from the obtained data. The findings suggest that the random forest clasifier performs best with the EEG of negative emotions, and investigates the effect of excluding individual and multiple electrodes from the EEG data on the system performance [135] ‘According toa study in [140], EEG can be used to detect neurophysiological changes associated with schizophrenia. In another study, external vestibular electrical stimulation was used to induce vertigo symptoms and EEG features were extracted using a wavelet decomposition algorithm. The extracted features were then classified into diferent levels of vertigo using logical regression, SVM, backpropagation, and RF clasifers. The RF model