Analysis of EEG signals and data acquisition methods: a review

ABSTRACT

Early illness diagnosis and prediction are important goals in healthcare in order to offer timely preventive measures. The best, least invasive, and most reliable way for identifying any neurological disorder is EEG analysis. If neurological disorders could somehow be predicted in advance, patients could be saved from their detrimental consequences. With promising new advancements in machine learning-based algorithms, Early and precise prediction might induce a radical shift. Here, we present a thorough analysis of cutting-edge AI methods for exploiting EEG data for Parkinson’s disease early warning symptoms detection, sleep apnoea, drowsiness, schizophrenia, motor imagery classification, and emotion recognition, among other conditions. All of the EEG signal analysis procedures used by different authors, such as hardware software data sets, channel, frequency, epoch, preprocessing, decomposition method, features, and classification, have been compared and analysed in detail. We will point out the difficulties, gaps and limitations in the current research and suggest future avenues of research.

KEYWORDS:

• EEG
• Parkinson
• Sleep apnoea
• Drowniness
• Schizophrenia

Introduction

Studies of biological processes in an organism, for example combinations of genes and proteins, heart and brain rhythmic patterns, and imaging of tissue and organs, are known to as bio signals. Biomedical signal processing aims to extract useful information from biological signals (Chang and Moura 2018). While electroencephalograms (EEG) are typically used to diagnose neurological diseases, electrocardiograms (ECG) are frequently utilised to diagnose cardiovascular disease. Biomedical images are used to investigate illnesses and are suggestive of anatomical information. There are a variety of commonly utilised modalities, ranging from ultrasound to radiology-based short-wavelength x-rays. For inexpensive and precise diagnostics, machine-based biomedical signal and image analysis and labelling are crucial. Biologists may learn new biological ideas through the application of biomedical signal processing, and doctors can keep track of specific disorders (Rajput et al. 2017). Numerous diagnostic and control applications based on the examination of bio signals. Biomedical signal processing’s primary objective in the past was to filter data to reduce noise (Majdalawieh et al. 2003). Classification is another new and crucial element in several biomedical research (Chang and Moura, 2018). Clinic classification is used to distinguish between pathology and normal. In clinics, classification is challenging for various reasons: It requires much time for experts to collect the necessary knowledge to accurately differentiate between comparable conditions, such normal and abnormal. Manually classifying each of these cases takes a lot of time and effort. The greatest difficulty arises when the signal characteristics are weak and thus difficult for professionals to distinguish[4]. Signal processing automatic categorisation algorithms have the potential to help overcome some of these challenges and support biological decision-making. An automatic classifier can identify features that are difficult to distinguish without bias, help human operators, and learn the categorical data from a database[4].

(Rasheed et al. 2021) An especially useful diagnostic technique for examining the brain’s signalling system throughout various psychological and physical processes is electroencephalography (EEG). Through EEG, a wide range of disorders’ prognosis and treatment have been extensively researched. The EEG signals evaluate the brain’s brainwave activity in a non-Gaussian and non-stationary manner, which is then utilised to diagnose the different types of brain illnesses. Brain function can be divided into normal and pathological with the aid of EEG analysis. Examining longer-duration EEG records is important for a precise diagnosis of epilepsy. It takes a lot of time and effort for experienced neurologists to analyse continuous EEG data that were recorded over a period of days, weeks, or perhaps even months. Machine learning (ML)-based prediction approaches have been employed in numerous studies to solve various problems over the years. A advanced machine learning (ML) technique is deep learning (DL), that can analyse enormous sets of data through a multilayer hierarchical design to learn patterns more precisely (Rasheed et al. 2021). Researchers have used Deep Learning techniques to address many real-world applications because of its capacity to deliver extremely accurate findings. A variety of researchers have proposed DL-based methodologies. In order to forecast various neurological conditions and anomalies, this study will discuss and expand on the key developments in EEG signal processing that have been made with the use of ML techniques. We will give a quick overview of neuroscience, the many technologies used to research the brain signal, and how they have been applied to EEG signal analysis.

The remainder of the article is organised as follows: Section II presents primary contribution, section III Background on EEG then presents review, analysis, comparison and key findings, the conclusion is presented in Section IV.

Primary contribution

Although there are various reviews, they all address only one neurological issue using EEG signals. There is no study that covers the application of EEG analysis for multiple disorders/conditions in depth. Rasheed et al., for instance, have published a detailed overview of models for epileptic seizures (Rasheed et al. 2021), Epileptic Spike Detection Algorithms for EEG and MEG were also examined by FATHI et al (Abd El-Samie et al. 2018), Seif Soliman et al. reviewed only for Deep Learning Approaches for Epileptic Seizure (Soliman et al. 2022), Mohammad-Parsa Hosseini et al reviewed Machine Learning for EEG Signal Processing for detection various neurological condition (Hosseini et al. 2021), An review of the development of seizure forecasting techniques from the 1970s to 2006 was presented by Mormann et al. (Mormann et al. 2007) and have addressed the key concerns with ES prediction approach. The outcomes of the prediction’s statistical significance have been thoroughly detailed by Gadhoumi et al. and gave a quick summary of the valid techniques employed for ES prediction (Gadhoumi et al. 2016). In the recently released review, Kuhlmann et al. noted briefly the improvements in ES prediction. They came to the conclusion that these improvements in regular statistical analyses are paving the way for the creation of ES prediction procedures, and they improved the current standards to bring about this development (Kuhlmann et al. 2018).

Apart from all the review that various author done till date, this survey document is remarkable since it offers detailed explanations. not only like classification but other preprocessing steps like resampling, noise removal, decomposition, feature selection etc. techniques for Parkinson, Sleep Apnoea, Drowsiness, Schizophrenia, Motor Imagery Classification and Emotion Recognition and explores prospective research directions along with research gap and challenges in this area.

Motivation

This article addresses a critical gap in current research by offering a comprehensive exploration of EEG signal analysis for multiple neurological conditions. While existing studies focus on singular issues, our article stands out by providing detailed insights into preprocessing techniques and classifications for Parkinson, Sleep Apnoea, Drowsiness, Schizophrenia, Motor Imagery Classification, and Emotion Recognition. Through this work, we aim to contribute valuable knowledge to the biomedical signal processing field, facilitating more accurate diagnostics and paving the way for future advancements in EEG-based machine learning applications.

Background on EEG for neuroscience and neuroimaging

EEG significance and diagnosed conditions

An EEG examination looks for irregularities in the brain’s electrical phenomenon. Electrodes will be positioned all over your head throughout the operation. Your brain activity generates very little electrical charges, which the electrodes detect. The electrical charges developed are shown as either a graph on a digital screen or a transcript drawn out on paper HealthHealth Home (Health home | Johns Hopkins Medicine) 2023. The EEG is used to detect multiple neurological disorders. When epilepsy is present, seizure activity will appear as acute waveforms on the EEG. Patients with brain lesions, which may be brought on by tumour tissues or clots, may have very slow EEG pulses depending on the size and place of the damage. This assessment might possibly be utilised to determine what makes the brain less active such as narcolepsy, various psychoses, Alzheimer’s disease, and others. The general electrical action of the brain may also be monitored utilising EEG (for instance, drug intoxication, assessing trauma, or determining the amount of brain damage in individuals who are unconscious.).The EEG is useful to monitor flow of blood inside brain during surgery. The characteristic of the raw EEG is Gamma(γ)higher than 30 Hz, Beta(β) (13 to 30 Hz), Alpha(α) (8 to 12 Hz), Theta(θ) (4 to 8 Hz), and Delta(δ) (less than 4 Hz) are the five frequency ranges. The following conditions are those that an EEG signal analysis may be used to identify ConditionsConditions an EEG ()2023

Seizure disorders

Epilepsy and other seizure disorders (Shoeibi et al. 2022) are the main conditions that EEG is utilised in assessment. A brain condition known as epilepsy is characterised by seizures, muscle spasms, and unusual behaviour or feelings. The exact cause of epilepsy is not fully understood by experts. It could entail atypical connections between nerve cells and chemical abnormalities in the brain. Medication to prevent seizures is the major treatment for epilepsy. Some seizure disorders are treated through brain surgery ConditionsConditions an EEG ()2023.

Sleep disorders

Narcolepsy to insomnia are examples of different sleep disorders (Sharma et al. 2022). Insomniacs have problems staying asleep. Anxiety, trauma, and particular illnesses or medications are the causes. Improved sleep hygiene and a change in lifestyle can assist. Excessive daytime fatigue and sleep episodes are symptoms of narcolepsy. Low amounts of the neurological component hypocretin may cause narcolepsy, according to experts. For the treatment of chronic insomnia and narcolepsy, doctors may prescribe medications Conditions an EEG ()2023.

Brain tumors

Brain tumours can take many different forms. They may be benign (not cancerous) or malignant (cancerous) (benign) (Raghavendra et al. 2023). Headaches, seizures, sensory changes, mobility issues, exhaustion, and altered behaviour are some of the signs of a brain tumour (Galgano et al. 2017). The majority of brain tumours have unknown causes, according to experts. Doctors use surgery, radiation, and chemotherapy to treat brain tumours. For most people, rehabilitation is essential to their recovery(Conditions an EEG 2023).

Brain injury

Severe neurological injuries can occur, which is caused by an outside force, or acquired, which is frequently caused by a stroke or a brain oxygen deficit. The seriousness of the injury will determine the symptoms, which may include everything from headaches and balance problems to comas and unconsciousness. Sleep, avoiding stress and physical activities like reading are part of the most basic treatments. Losing mental ability necessitates longer term therapy. In the case of moderate and severe brain injuries, lifelong therapy may be required Conditions an EEG ()2023.

Dementia

Memory loss and altered mental abilities, such as speaking and understanding, are symptoms of dementia. Behavioural issues can also happen. Vascular dementia, Alzheimer’s condition (Khare and Acharya 2023a), are just a few of the illnesses that can cause dementia. Medication is used by doctors to treat dementia in order to improve cognitive function and memory. They also employ methods for dealing with behavioural issues. But medication cannot undo the alterations in the brain that are developing dementia Conditions an EEG ()2023.

Brain infections

Brain abscesses and encephalitis, an infection-related brain inflammation, are examples of disorders that can result from brain infections. Headaches, seizures, mental state changes, and behavioural or personality abnormalities can all be brought on by brain infections. Typically, an infection travels from another region of the body to the brain. Surgical procedure and medication if required are used by doctors to treat brain infections Conditions an EEG ()2023.

Stroke

Once the mind’s blood circulation is interrupted, a stroke happens and destroy brain cells. Sudden weakness or paralysis, numbness, a sagging face, alterations in speech, or difficulty interpreting words are all indications of a stroke. A blood artery begins to bleed blood and enters the brain, or a Blockage of a blood vessel, causing an ischaemic stroke (haemorrhagic stroke). Emergency care is necessary for stroke Conditions an EEG ()2023.

Attention disorders

Problems involving lack of attention are attention deficit disorder (ADHD) and attention deficit hyperactivity disorder (ADHD) (Khare et al. 2022; Khare and Acharya 2023b). They are characterised by difficulties maintaining focus, giving proper attention, and managing activity. Attention difficulties are caused by unknown factors, according to experts. Genes might be involved. Doctors use medication, behavioural therapy, and social skills instruction to address attention issues Conditions an EEG ()2023.

Behavior disorders

Conduct disorder (CD) and Oppositional defiant disorder (ODD) are examples of abnormalities of behaviour (Khare et al. 2023, 2023). If these illnesses have a physical origin, an EEG can assist in identifying it. Behaviour disorders impair connections in the workplace, at home, in the family, and in the community. Behaviour modification therapy is beneficial for those who have these problems. In some circumstances, physicians also prescribe drugs Conditions an EEG ()2023.

Developmental delays

Communication, reasoning, and motor skills are all impacted by developmental delays (Khare et al. 2023). Children with developmental delays lose or miss out on typical developmental milestones like walking and talking. These kids’ delays may have a physical cause, which an EEG may help find. The focus of treatment is on a child’s particular requirements. This may involve social skills instruction, behavioural/physical/occupational/speech therapy Conditions an EEG ()2023.

Parkinson’s disease

Parkinson’s disease (Parkinson’s disease - symptoms and causes - mayo clinic (2023) Movement is impacted by Parkinson’s disease (Khare et al. 2021b), a debilitating condition of the neurological system. Early signs might include a little shaking in just one hand. Although the illness usually causes tremors, it also frequently causes movement to slow down or stiffen. Parkinson’s symptoms and indicators include Parkinson’s disease (Parkinson’s disease - symptoms and causes - mayo clinic (2023): Numbness, sluggish motion (bradykinesia), stiff muscles, poor balance and posture, a reduction in automatic motions, Speech abnormalities, Writing variations.

Phases of EEG analysis

Broadly Phase of Biomedical signal processing can be divided into following phases and details of these steps are given below

Signal recording

(Penzel et al. 2006)In several areas of contemporary medical services recently, over a few decades, tracking and recording bio-signals have become crucial. This shows that the significance of physiological control systems is acknowledged. ECG is the most well-known use of bio signal monitoring (ECG). The ECG has undoubtedly had the longest history in bio signal monitoring and recording due to its robustness and strength (amplitude approximately 1 mV). For a thorough diagnosis of particular heart issues, a cardiologist may record it with 6 or 12 leads. Bio signals are now routinely recorded during diagnostic procedures in neurology/neurophysiology (EEG, EMG) and cardiology (ECG, blood pressure). Several signals have an electric component and are electrophysiological in origin (e.g. ECG, EEG). To extract these signals with the fewest artefacts possible, careful amplification and filtering are required.

• Radio signals and powerful magnetic forces are used in magnetic resonance imaging (MRI) to provide detailed images of the interior of the body.

• The technique of magnetoencephalography (MEG) records the beams of magnetism created by the electrical ions that typically flow in the brain in order to map brain function.

• An electroencephalogram (EEG) is a scan that uses fine wire chips (electrodes) connected to head to find electrical movement in mind.

• At specialist medical facilities, data from human intracranial electroencephalography (iEEG) is captured using electrodes positioned on or implanted in the brain.

• ECoG can be carried out intraoperatively (during surgery) or extra operatively (during non-surgery) (extra operative ECoG).

• During an MRI imaging process, blood that is entering the body is magnetically labelled called as arterial spin labelling (ASL) allows for an evaluation of blood flow to the brain without surgery.

• Among the nuclear medicine techniques is positron emission tomography (PET), keeps track of the physiological function of the cells in diverse bodily tissues. In actuality, PET merges biochemical analysis with nuclear medicine [Figure 1].

Figure 1. Wide range of physiological characteristics in various medical contexts.

EEG signal recording

(Rasheed et al. 2021)Based on where the reference electrode is placed, there are two different ways to capture the EEG:

1. Bipolar Montage: The voltage differential between the electrodes in a bipolar montage is observed while both electrodes are applied to a scalp area that is electrically active.

2. Monopolar Montage/Unipolar Montage: One active electrode and one or more coupled reference electrodes make up a monopolar montage. Electrical neutrality should be maintained for the reference electrode and so is feasible when compared to brain activity. The recorded signals serve as a representation of the differential between the reference electrodes and the active brain regions. The top of the nose, the left or right mastoid, the chest, and the non-cephalic, balanced stern vertebral lead are common reference points.

Following are the various commercial hardware available for EEG recording:

• OpenBCI

• g.USBamp amplifiers

• EMOTIV

• QUASAR

• BioSemi

• Axxonet

Software used in EEG processing

(Holdgraf et al. 2019) A guideline for representing the data in neuroscience and a set of programs and tools that make it easier to use BIDS data structures are the two main outputs of the community project BIDS. In order to create the iEEG-BIDS specification, various new community tools had to be developed. We have listed those who needed the most recent iEEG-BIDS development below.

• The BIDS validator: - The BIDS specification is tested using a validator (written in JavaScript). It checks text files for the necessary file types and information required by the BIDS standard and parses a folder to see if its naming conventions and hierarchy are compliant with BIDS. An extra set of data fields and directory structures were added by the iEEG-BIDS addition so the BIDS validator had to be changed to support these.

• The BIDS starter kit: -is a set of wiki resources, tutorials with stakeholder input, utility programs, and instructions aimed at introducing researchers to the BIDS data format. These materials include the two main programming languages that iEEG-BIDS supports (Matlab and Python).

• MNE-BIDS and MNE-Python. A Python open-source application for data visualisation, analysis, and electrophysiological. The MNE-Python package just received an update to add support for iEEG. The group collaborated with the MEG and EEG teams to create MNE-BIDS, a updated package for data transmission among MNE-Python and the BIDS requirements in electrophysiology, additionally to the development of the iEEG-BIDS standard. Electrophysiology procedures in MNE-Python can be easily converted to the BIDS dataset framework with the help of MNE-BIDS.

• FieldTrip. In the Matlab programming language, FieldTrip is a free toolkit for studying electrophysiology. It can export data to BIDS specified forms and enables importing data from a wide range of iEEG formats. Additionally, FieldTrip has a method called data2bids.m that enables users to properly annotate their iEEG and MRI data and arrange it.

• EEGLAB is a GUI-based toolkit for handling MEG, EEG, and other electrophysiological data that also contains independent component analysis (ICA), Fourier/wavelet assessment, component removal, statistical data relating to events, and a range of practical methods to analyse separate and averaged data.

• iElvis:-An open-source Matlab application called iELVis is used to locate iEEG electrodes and display their data layer on neuroimaging. Using BIDS-iEEG protocols, it would acquire and output iEEG electrode locations and related neuroimaging data.

• Brainstorm. An open-source Matlab program called Brainstorm is used to analyse multimodal neurophysiology data.It offers interfaces for users for setting up electrode connections in clinical iEEG signals, post-implant CT or T1 scans, as well as analysing and editing huge information. Most iEEG file types are supported, and they may be converted into BrainVision/EDF files. From the Brainstorm database, studies may be automatically loaded and generated as data structured for BIDS.

• OpenNeuro. A database for freely accessible neuroimaging records is Open Neuro (OpenNeuro.org). The bids-validator, which heavily relies on the BIDS standard, is used to validate each dataset before to submission. This facilitates the online sharing and storage of EEG datasets for researchers who employed BIDS in their labs.

Dataset and brain imaging data structure (bids) format

ECoG, ASL, PET, MRI, MEG, EEG, and iEEG data can be shared on the free and open platform known as OpenNeuro. Complex data from neuroimaging research can be structured in a variety of ways. There is currently no agreement on how to distribute and organise the data gathered from neuroimaging studies. It is possible for two researchers working in the same lab to choose to organise their data differently. Lack of agreement (or a standard) causes misunderstandings, which wastes time rearranging data or rewriting scripts that were written with a specific structure in mind. Here, we outline an easy-to-use system for organising behavioural and neuroimaging data. You will gain by applying this standard in the following ways:

• Your data will be simple for another researcher to work with. You simply need to refer to this document to understand how the files are organised and formatted. This is crucial if you run your own lab and expect to have multiple people working on the same data at once. By utilising BIDS, you can avoid wasting time attempting to decipher and reuse data that has already been collected by a graduate student or postdoc who has left the lab.

• Software tools for data analysis are becoming more and more prevalent, and many of them can comprehend BIDS-organised data.

• Datasets organised using BIDS are accepted by databases like OpenNeuro.org. By utilising BIDS to structure and characterise your data as soon as it is acquired, you can reduce the extra time and effort required for publishing and simplify the curation process if you ever intend to distribute your data publicly (these days, several journals mandate this).

• You may quickly identify missing values by using validation tools like the BIDS Validator to evaluate the consistency of your dataset [Figure 2].

Figure 2. BIDS format structure with detail info inside (Pernet et al. 2019).

Signal preprocessing for noise/artifact removal

(Repovš 2010)– (Sharma et al. 2022) Many methods are available for removing the artefact/noise from EEG signal like Common Spatial Patterns (CSP), Single Value Decomposition (SVD), Normalised Least Mean Square with adaptive filter, Independent Component Analysis (ICA), Common Average Referencing (CAR), Filtering, Surface Laplacian (SL), Principal Component Analysis (PCA) etc. After removing noise/artefact, signal analysis results more accurate report

Signal decomposition

Retrieval and isolation of signal components from complex signals are the objectives of signal decomposition. Following are the various recently updated method for signal decomposition:

• EEG signals are used in empirical mode decomposition (EMD), which allows for meaningful time-frequency-amplitude analysis of the signals. The suggested method applies the EMD algorithm to decompose EEG signals into AM-FM components (referred to as IMFs).

• Flexible variational mode decomposition (F-VMD) can be used for the adaptive breakdown of signals.

• Short-time Fourier transform (STFT) and Fast Fourier transforms (FFT) for the breakdown of signals.

• Higuchi algorithm

• The non-stationary EEG signals are decomposed using optimal variational mode decomposition (O-VMD) and the eigenvector centrality method (EVCM), which chooses a prominent channel.

• Minimum average maximum tree (MAMT)

• adaptive variational mode decomposition (AVMD),

• Tunable Q wavelet transforms (TQWT)

• Flexible analytic wavelet transforms (FAWT)

• Singular value decomposition (SVD),

• Empirical wavelet transforms (EWT) and empirical wavelet packet transform (EWPT),

• Automated tunable Q wavelet transform.

• Principle component analysis (PCA) and independent component analysis (ICA) for decomposing signal into another dimension.

• Optimised TQWT (O-TQWT),

• Teager energy operator and empirical mode decomposition (EMD),

• Non-chirp variational mode decomposition (NCVMD),

• Smoothed Pseudo Wigner Ville Distribution

Signal feature extraction

(Subasi 2019; Siuly et al. 2020; Rasheed et al. 2021) To optimise effectiveness for the categorisation of biological data, feature identification is necessary. The various features are

1. Activity natural,

2. Adaptive hermite decomposition (AHD),

3. Complexity,

4. Crest factor (CF)

5. Difference absolute standard deviation value (DASDV):

6. Entropy of approximation (APPEN),

7. Entropy.

8. Fano factor,

9. 1st quartile (q1), 2nd quartile (q2), 3rd quartile (q3), interquartile range (iqr)

10. Higuchi fractal dimension (HFD),

11. Sure entropy,

12. Hurst exponent (HE),

13. Interquartile range

14. Kurtosis,

15. Lampel-ziv complexity (LZC) feature,

16. Hjorth complexity (HJC) and mobility (HM)

17. Log energy (LEEN) entropy,

18. Mean,

19. Median,

20. Min, max, sum, first, last

21. Mobility,

22. Mode,

23. P10, p25, p50, p75, p90,

24. Permutation (PREN) entropy,

25. Renyi(REEN) entropy

26. Root mean square,

27. Sample entropy and conditional entropy.

28. Second moment,

29. Shannon entropy, permutation entropy,

30. Simple square integral (SSI),

31. Skewness,

32. Spectral flatness (SF),

33. Standard deviation,

34. Trimean,

35. Tsallis (TSEN) entropy

36. Variance,

Feature ranking

Selecting the top significant attributes of a particular feature set is the process of feature selection. Both classification and feature selection can improve a model’s performance (Siuly et al. 2020; Khare and Bajaj 2020; Murugappan et al. 2020) Discrimination ability of the considered features calculated by following methods

1. Chi Squared: By determining the chi-squared value with respect to the class attribute for each attribute, one can determine the importance of a feature.

2. Correlation: returns the value’s normalised or absolute form as the weight after calculating the correlation between each attribute and the label attribute.

3. Gini Index: to determine the significance of the attributes.

4. Information Gain: Considering the information gained, calculates the importance of the attributes.

5. Kruskal Wallis test: to pick the most discriminating features while reducing the dimensionality.

6. Maximum Relevance: Depending on the feature and label type (numerical/nominal), picks mutual information, Pearson correlation or F- test.

7. PCA: uses one of the primary components’ elements as the basis for the feature weights.

8. Relief Measures

9. SAM (Significance Analysis for Microarrays)

10. SVM: uses the coefficients of a linear support vector machine’s normal vector as the attribute strengths.

11. Uncertainty: measures the symmetrical uncertainty with regard to the class to determine the significance of a characteristic.

12. Univariate post-hoc analysis: The post-hoc analysis approach is a non-parameterised way to determine a feature’s capacity for discrimination.

Classification

Classification involves identifying, comprehending, and organising signals and feature sets into predetermined categories or subpopulations (Subasi 2019; Rasheed et al. 2021; Yaacob et al. 2023; Khare et al. 2024) Followings are the various recently used classification methods (Tables 3 and 4):

1. ANN

2. Bagged tree (BaT) classifier

3. CNN architectures AlexNet, VGG16, ResNet50, Inception

4. Decision tree (DT), basic tree, medium tree and complex tree are the three kernels (El-Dahshan et al. 2024).

5. Discrete wavelet transforms (DWT)

6. The two types of discriminant analysis are linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA)

7. Dynamic graph CNN

8. Ensemble learning: By integrating many classification models, it is utilised to increase the prediction performance of the constituent model. Subspace discriminant, bagged trees, RUS Boosted trees, and subspace KNN are among the ensemble learning approaches.

9. Extreme learning machine (ELM) kernels are sinusoidal (Sin), radial basis function (RBF), triangular (TB) and sigmoid (Sig).

10. Flexible extreme learning machine (F-ELM)

11. Six kernels: fine, medium, coarse, cosine, cubic, and weighted in K-nearest neighbours (k-NN)

12. Support vector machine (SVM)

13. Least square SVM (LSSVM)

14. Linear discriminant analysis (LDA).

15. Linear regression (LR)

16. Using the fuzzy neural network and the moving average power spectrum

17. Multi-layer perceptron (MLP) classifiers

18. Naïve Bayes (NB)

19. Optimal centre constructive covering algorithm (OCCA)

20. Pooling-based deep recurrent neural network(PDRNN)

21. Probabilistic neural network (PNN)

22. Random forest classifiers

23. Recurrent neural networks (RNN).

Performance measuring parameter calculation

The most important task in building any classification model is to evaluate its performance (Taran and Bajaj 2019a; Murugappan et al. 2020; Taran et al. 2021)Following metrics evaluates the performance of the proposed models on Table 1.

• Accuracy

• Cohen’s kappa

• F1-score

• False alarm rate

• Fowlkes-mallows index (FM)

• Mathew’s correlation coefficient (MCC)

• Precision

• Receiver operating characteristic (ROC) curve.

• Sensitivity

• Specificity.

Current state of art

Analysis and comparison: detection of Parkinson’s disease

Table 1. comparison of parkinson’s disease detection methods

Authors	Data set	Channel, Frequency	Preprocessing	Epoch	Decomposition Method	Features	Classification	Accuracy	Year
Jeong et al.41	EEG of 26 PD-dementia, 26 Alzheimer’s disease and the 26 healthy subjects	21,200Hz			Wavelet Transformation	FR thresholds	LDA	81	2016
Liu et al.42	EEG data 42 participants	10, 250Hz	DWT then compute entropy		three level DWT (Discrete Wavelet Transformation)		optimal center constructive covering algorithm (O CCA)	90	2017
Chaturvedi et al.43	EEG of 50 PD patients	256,500Hz	independent component analysis		frequency decomposition		RF, SVM, RANDOM FOREST	72.4	2017
Taoufiq et al.44	Speech	14,128Hz	wavelet transform then extracting Mel Frequency Cepstral Coefficients (MFCC)		Wavelet Transformation	MFCC	SVM	86.84	2019
Ruffini et al.45	EEG idiopathic RBD patients(121) and healthy controls(91)	14,60Hz	Fourier analysis (FFT) with a Hann window	20 second 14 channel	Fourier analysis (FFT)		RNN	85	2019
Tuncer et al.46	756 signal, 252 people said ‘‘a’’ vowel	NS, 44.1 KHz	MAMa tree to generate Q4 novel feature		singular value decomposition (SVD)	The SVD used in feature extraction	linear discriminant (LD), SVM	96.83	2020
Anjum et al.47	EEG signal of (27+14) PD patients & (27+14) controls	62,500Hz			linear predictive coding	late positive complex (LPC) to capture PSD	principal component analysis	85.5	2020
Oh et al.48	20 PD patient’s EEG signals	14,128 Hz	following the threshold method, a 6th-order bandpass Butterworth filter	1588 artifact-free epochs	not specified		CNN, DL	88.25	2020
Xu et al. 49	20 PD patient’s EEG signals	14,128 Hz	following the threshold method, a 6th-order bandpass Butterworth filter		not specified		FKNN	88.31	2020
Smith et al.50	EEG dataset of 15 PD and 16 healthy controls	32,512 Hz		not defined	TQWT. The wavelet transforms (WT) LPS and HPS filter banks	HJC,Minima, HE,HM, and HFD	ANN, SVM, LSSVM, KNN, RF	97.65	2021

Figure 3. Chart describing Parkinson’s disease detection methods.

Detecting schizophrenia is a complex process that often involves a combination of clinical evaluation, behavioral observations, and various assessment tools. It’s important to note that only qualified healthcare professionals, such as psychiatrists, psychologists, and other mental health specialists, should diagnose schizophrenia [Figure 3]. However, researchers and clinicians use different methods for early detection and assessment [Table 2].

Analysis and comparison: detection of schizophrenia

Table 2. Comparison of schizophrenia detection methods.

Authors	Data Set	Channel, Frequency	Epoch	Preprocessing	Decomposition	Features Extraction	Classification	Accuracy	Year
Sabeti et al. (2007).	self-recorded	22,200 hz	30 sec	Average filter and Butterworth band pass filters (order 5)	DWT	Autoregression, band power, andwavelet energy	In order to classify the two group’s condensed feature sets, LDA and SVM are used.	88	2007
Boostani et al. (2009).	self-recorded	22,100 hz	32 sec	Butterworth filter, order five		Band power, fractal dimension, and the variables of the autoregressive (AR) model	support vector machine (SVM), Adaboost, fuzzy SVM (FSVM),and LDA	50	2009
(Sui et al. 2014).	self-recorded	64,1000 hz	60 sec	filtered by a band-pass (.01–.5 Hz)	DFT	Multi-set canonical correlation analysis	MCCA, SVM RFE	91	2014
Kim et al. (2015).	self-recorded	64,1000 hz		a 1-Hz high-pass, a 60-Hz notch filter and ICA	Fourier Decomposition	Five-frequency study of the spectra	ANOVA	62	2015
Johannesenet al. (2016).	self-recorded	64,1024 hz	32 sec	baseline, encoding, retention, and retrieval stages of t4 processing	Spectral Decompositon	5 frequency components: theta1, theta2, alpha, beta, gamma	SVM models	84	2016
Shim et al. (2016).	self-recorded	64, 1000 hz	100 ms	.1–100-Hz band-pass and 60 Hz notch filter.	Spectral Decompositon	source-level features as well as sensor-level features	support vector machine (SVM) classifier with leave-one-out cross-validation (LOOCV)	88.24	2016
Piryatinskaet al. (2017).	http//brain.bio.msu.ru/eeg_schizophrenia.htm	16,128 hz	30 sec	Bandpass Filter	ε-complexity	ϵ-complexity coefficients	RF,SVM	85.3	2017
Devia et al. (2019).	self-recorded	32,40 hz		a bandpass filter (Butterworth)	event-related potentials (ERP)	Event-related potentials (ERPs) features	LDA, Combination of Two LDA	71	2019
Zhang (2019)	self-recorded	64,1024 hz	3 sec	bandpass filter .5 Hz and 15 Hz.	not defined	Event related potential	Random forest	80	2019
Jahmunahet al. (2019).	self-recorded	19,250 hz		encoding, retention, and retrieval during the pre-stimulus baseline	not defined	t-test assisted variable selection and non-linear features extraction	Probabilistic-Neural Network (PNN),Decision-Tree (DT), k-Nearest-Neighbour (KNN), Support-Vector-Machine (SVM), Linear-Discriminant analysis (LD),	92.91	2019
Varun et al. (Siuly andZhang, 2016)	self-recorded	72,1024 hz		encoding, retention, and retrieval during the pre-stimulus baseline	empirical mode decomposition (EMD)	Twenty-two statistical characteristics are derived from IMFs using the EMDalgorithm.	Decision tree,Ensemble Bagged tree, K-nearest neighbour,Support vector machine.	89.5	2020
(Sharma and Acharya 2021).	self-recorded	32,250 hz	25 sec	encoding, retention, and retrieval during the pre-stimulus baseline	Wavelet Decomposition	optimal wavelet-based II-norm	KNN and EBT	99.21	2020
(Khare et al. 2021c).	Kaggle	32,250 hz		bandpass filter .5 Hz and 15 Hz.	EWT	Smoothed pseudo-Wigner Villedistribution	Alex Net, VGG16, and ResNet50 networks	93.36	2021

Figure 4. Chart describing schizophrenia detection methods.Chart describing motor imagery classification technique is discribe in Figure 4.

Analysis and comparison: motor imagery classification

Table 3. Comparison of motor imagery classification technique.

Authors	Data Set	Channel, Frequency	Epoch	Preprocessing	Decomposition	Features	Classification	Accuracy	Year
(Ince et al. 2009).	IVa of BCI Competition III	118,1000		low pass filtered and down sampled to 125 Hz	empirical mode decomposition (EMD)	8	PCA, LDA	96	2009
Siuly et al. (2011).	IVa of BCI Competition III	118,1000		Band-pass filter .53–40 Hz	Clustering technique (CT)	9	clustering technique-based LSSVM	93.66	2010
(Xu et al. 2020).	IVa of BCI Competition III	118,1000			empirical mode decomposition (EMD)	6	RCSP for signal decomposition	83.9	2010
(Siuly and Li 2012).	III BCI Competition IVa and IVb	102 &118,1000		around .05 and 200 Hz band-pass filtered	cross-correlation (CC)	–	LS-SVM	95.18	2012
Zhang et al. (2013).	IVa of BCI Competition III	118,1000		around .05 and 200 Hz band-pass filtered		6	NDA, LIBSVM, Z-LDA to LDA, SVM, and HLDA.	98.16	2013
Siuly et al. (2013).	III BCI Competition IVa and IVb	102 &118,1000			cross-correlation (CC)	6	LS-SVM	91.79	2013
(Siuly and Li 2015).	III BCI Competition IVa and IVb	101 &118,1000		around .05 and 200 Hz band-pass filtered	optimal allocated technique	11	Naïve Bayes (NB)	96.62	2014
Siuly et al. (2014).	III BCI Competition IVa and IVb	102 &118,1000		around .05 and 200 Hz band-pass filtered	cross-correlation (CC)	6	cross-correlation based logistic regression algorithm	94.23	2014
(Siuly and Zhang 2016).	III BCI Competition IVa and IVb	100 &118,1000			optimal allocated technique	11	Naïve Bayes (NB) probabilistic mode	91.97	2015
Verma et al. (2014).	IVa of BCI Competition III	118,1000			DWT (Akbari et al. 2023),(Akbari et al. 2021)	9	Least-square support vector machine (LS-SVM), logistic regression (LR), multilayer perceptron neural network (MLP), probabilistic neural network (PNN) and kernelized logistic regression (KLR)	94.9	2015
Taran et al. (2018).	BCI competition III	118,1000		Spatial filter based on sparsity	Hilbert transform and empirical mode decomposition (EMD)	5	LSSVM	96.89	2017
(Kevric and Subasi 2017).	BCI competition III	118,1000		around .05 and 200 Hz band-pass filtered. sampling frequency down-sampled from 1000 to 100 Hz	Wavelet packet decomposition (WPD), Empirical mode decomposition (EMD)	6	KNN	92.8	2017
Chaudhary et al. (2019).	BCI competition III	118,1000		Morse wavelet is used to produce a CWT filter bank.	Wavelet packet decomposition	–	DCNN (Sadiq et al. 2022)	99.33	2018
(Taran and Bajaj 2019b).	BCI competition III	118,1000	2000 samples	Input to the second-level filter bank is given by the first-level lowpass sub-band filter.	TQWT	4	least-squares support vector machine	96.85	2018
Park et al (Park et al. 2018).	BCI competition III	118,1000		filtered through a bank of 5-order Butterworth bandpass filters with 9 distinct frequency bands. EEG is projected into the feature space using R-CSP.	FRCSP	2	Ensemble Classification	82.59	2018
Khare et al (Khare and Bajaj 2020).	BCI competition III	118,1000	1000 samples	Butter-worth filter	flexible VMD	2	artificial neural networks (ANN), SVM, random forest classifier (RFC), and k-nearest neighbour (k-NN	99.35	2020
Chaudhary et al (Chaudhary et al. 2020).	BCI competition III	118,1000		right hand and foot applying CSP	FAWT	3	LDA is employed in the classification process. Few Laplacian EEG channels and F-score of the Fisher discriminant kind	95.17	2020
Xiaojun Yu et al (Yu et al. 2021)	BCI competitions II and III	118,1000	280 samples	multiscale principal component analysis (Sadiq et al. 2020)	empirical Fourier decomposition	10	feedforward neural network	88	2021
Muhammad et al (Sadiq et al. 2021)	BCI competitions II and III (Sadiq et al. 2022) (Sadiq et al. 2019),	118,1000	280 samples						2021

Figure 5. Chart describing motor imagery classification technique.

Chart describing motor imagery classification technique is discribed in Figure 5.

Analysis and comparison: emotion recognition

Table 4. Comparison of emotion recognition models.

Authors	Data Set	Channel, Frequency	Epoch	Preprocessing	Decomposition	Classifier	Accuracy	Year
Murugappan et al. (2010).	self-recorded	64,256		Surface Laplacian (SL) filter	Wavelet Transform	kNN	91.67%	2010
Wang et al. (2011).	self-recorded	64,1000	1000-point		FFT	SVM	66.51%	2011
Lin et al. (2010).	self-recorded	32, 500	16 30s segments		short-time Fourier transform (STFT)	SVM	82.29%	2010
Bajaj et al. (2018).	self-recorded	24,256			FAWT	weighted-kNN	96.12%	2018
(Krishna et al. 2019).	self-recorded	24,256			TQWT	ELM	87.10%	2018
QungQW et al. (2018).	received from the Penang General Hospital’s Neurology Department	9 100 hz	15–25 s		EWT andEWPT	KNN, PNN, and ELM using the RBF kernel	95%	2018
Taran et al. (2019a).	self-recorded				EMD-VMD	MC-LS-SVM	92.79%	2019
(Khare and Bajaj 2021c).	self-recorded	24,256			SPWVD	CNN	93.01	2020
(Murugappan et al. 2020).	self-recorded	14,128		Butterworth filter and IIR notch filter	ATQWT	MC-LS-SVM	96.16%	2020
(Khare and Bajaj 2021b).	self-recorded	24,256			O-VMD	ELM	97.24%	2020
(Khare and Bajaj 2021b).	self-recorded	24,256			O-VMD	kNN	93.80%	2020
(Khare and Bajaj 2021b).	self-recorded	24,256			O-VMD	SVM	93.10%	2020
(Khare and Bajaj 2021b).	self-recorded	24,256			O-VMD	DT	87.70%	2020
deOliveira et al. (2020).	self-recorded	61 512 hz	10s	60 Hz notch FIR filter, .5 to 50 Hz bandpass filter	Partial directedcoherence (PDC)	random forest	99%	2020

Figure 6. Chart describing emotion recognition models.Chart describing emotion recognition models was discribe in Figure 6 and Table 5.

Analysis and comparison: the sleep apnoea detection

Table 5. Comparison of sleep apnoea detection models.

Authors	Data Set	Channel, Frequency	Epoch	Preprocessing	Decomposition	Classifier	Accuracy	Year
Lin et al. (2006).	MIT/BIH sleep database	250 hz		multi-resolution wavelet transforms	Wavelet transform and wavelet coefficients features	ANN	69.64	2006
(Lee et al. 2012).	MIT/BIH sleep database	250 hz		The scaling function in the DWT and the Daubechies wavelet (db3) were employed in this work.	GreyART networks and wavelet coefficients features	GreyART network	83.13	2011
Zhou et al. (2015).	Two PhysioNet Databases	not mentioned	each 30 s, total 720 values	Detrended fluctuation analysis		SVM	95	2015
Almuhammadi et al. (2015).	MIT/BIH polysomnographic database	C3- O1, C4-A1 and O2-A1 channels and 250 Hz	30 sec time	Frequency domain	Energy, Variance	SVM, ANN, LDA and NB	97.14	2015
Shahnaz et al. (2016).	MIT/BIH polysomnographic database	a single channel (C3-A2 or C4-A1)	30 sec time epoch	Wavelet Smoothing	Delta band power ratio	KNN and SVM	84.07	2016
Taran et al. (2017b).	MIT/BIH polysomnographic database	256 Hz	30 sec, total normal 1195 and apnoea 947	TEO, Empirical Mode Decomposition	Features of TEO and AM-FM components	LS-SVM, KNN	95.29	2017
Taran et al. (2017a).	MIT/BIH polysomnographic database	256 Hz	30 sec, total normal 1195 and apnoea 948	Hermite decomposition algorithm	Features of HD and four Hermite coefficients	PSO-LSSVM	98.8	2017
(Bhattacharjee et al. 2019).	PhysioNet Databases		30 sec	Hermite decomposition algorithm	The Rician model and statistical characteristics	KNN	97.4	2019
(Taran et al. 2021).	MIT/BIH polysomnographic database	256 Hz	30 sec, total normal 1195, apnoea 948	TQWT and LZC	TQWT and LZC feature	Subspace-KNN, Decision Tree, Ensemble Classification Model	96	2021

Figure 7. Chart describing sleep apnoea detection models.Chart describing sleep apnoea detection models was discribe in Figure 7.

Analysis and comparison: drowsiness detection

Table 6. Comparison of drowsiness detection models

Authors	Data Set	Channel, Frequency	Epoch	Preprocessing	Decomposition	Features	Classification	Year	Accuracy
Correa et al (Garces Correa and Laciar Leber 2010).	MIT/BIH Polysomnographic EEG database	7,250 Hz	30 sec	2nd order, bidirectional, Butterworth, band – pass filter	PSD and Wavelet	12	Levenberg-Marquardt backpropagation algorithm NN	2010	86.5
Correa et al (Garcés Correa et al. 2014).	MIT/BIH Polysomnographic EEG database	250 Hz	5 sec	2nd order, bidirectional, Butterworth, band – pass filter	Wavelet Decomposition	7	ANN	2014	87
Boonnakelal (Boonnak et al. 2015).	MIT-BIH Polysomnographicdatabase	7,250 Hz		Energy coefficients with probability distribution (ECPD)	Wavelet	7	ANN	2015	90.27
Tripathy et al (Tripathy and Rajendra Acharya 2018).	MIT/BIH Polysomnographic EEG database	7,250 Hz		iterative filtering (IF) and Band pass filtering	not defined	10	Based on 10-fold cross-validation, DNN classifier	2018	91.33
Taran et al (Taran and Bajaj 2018).	MIT/BIH Polysomnographic EEG database	7,250 Hz	30 sec total 667 and 488		Hermite Decomposition	4	ELM (hidden layer with feed forward)	2018	95.45
Anita et al (Anitha 2019).	Yawning Detection Database and MIT-BIH Polysomnographic database	7,250 Hz	30s total 7500		not defined	8	SVM	2019	86.7
Budak et al (Budak et al. 2019).	MIT/BIH Polysomnographic EEG database	3,250 Hz	30 sec		TQWT		long-short term memory (LSTM) classifiers, AlexNet and VGG16	2019	94.31
Bajaj et al (Bajaj et al. 2020).	MIT-BIH Polysomnographicdatabase	7,250 Hz			TQWT		DT, LR, fine Gaussian support vector machine, weighted KNN, boosted trees and extreme learning machine (ELM)	2020	91.84
Khareet al (Khare andBajaj 2022).	MIT/BIH Polysomnographic EEG database	7,250 Hz	660 alert and 480 drowsy EEGepochs		0-TQWT	5	LS-SVM	2021	96.14
Bajaj et al (Sharma et al. 2021).	MIT/BIH Polysomnographic EEG database	250 Hz	30 sec	Flexible analytic wavelet transforms	EMD (EmpiricalMode Decomposition)	5	ELM	2021	95.6
Varun et al (Khare andBajaj 2021a).	MIT/BIH Polysomnographic EEG database	7,250 Hz	30 sec total 667 and 488	AVMD	Adaptive Variational ModeDecomposition	5	ANN	2021	97.19

Figure 8. Chart describing drowsiness detection models.

Key Findings

EEG Signal Recording Hardware Type

By detail analysis of papers reviewed in this study it has been found the BioSemi active two system (Abd El-Samie et al. 2018; Khare et al. 2021b; Khare and Acharya 2023a, 2023b) were used in most of the study [Table 6 and Figure 8] .

Software Type

Most of the work used MATLAB to implement their methodology of the research work.

Dataset Type

Following are the dataset used in various study:

• MIT/BIH polysomnographic database

• Self-recorded

• brain.bio.msu.ru/eeg_schizophrenia.htm

• III BCI Competition IVa and IVb

• Collected from Neurology Department of the Penang General Hospital’s [Table 7]

Channel And Frequency Of The Dataset.

following table represent the list of various channel & frequency used by no of authors

Table 7. Number of channel & frequency count

AUTHORS	DATA SET	CHANNEL, FREQUENCY
1	10,250Hz	1
2	100&118,1000Hz	1
3	101&118,1000Hz	1
4	102&118,1000Hz	1
5	103&118,1000Hz	1
6	104&118,1000Hz	1
7	118,1000Hz	13
8	14,128Hz	4
9	14,60Hz	1
10	16,128Hz	1
11	19,250Hz	1
12	21,200Hz	1
13	22,100Hz	1
14	22,200Hz	1
15	24,256Hz	7
16	24,256Hz	7
17	256,500Hz	1
18	3,250Hz	1
19	32,250Hz	2
20	32,40Hz	1
21	32,500Hz	1
22	32,512Hz	1
23	61,512Hz	1
24	62,500Hz	1
25	64,1000Hz	4
26	64,256Hz	1
27	7,250Hz	8
28	72,1024Hz	1
29	9,100Hz	1

Figure 9. Chart representing count of channel & frequency used.

Signal Preprocessing

Following are various methods used by different author for pre-processing [Figure 9]

• threshold technique Then, a 6th-order bandpass Butterworth filter

• 2nd order, bidirectional, Butterworth, band – pass filter”

• iterative filtering (IF) and Band pass filtering

• Butterworth band pass filters (order five) and average filter “Butterworth filter, order five

• band pass filtered (0.01- 0.5 Hz) ”1-Hz high-pass filter and a 60-Hz notch filter

• low pass filtered and down sampled to 125 Hz”

• Band-pass filter 0.53–40 Hz

• between 0.05 and 200 Hz band-pass filter

• Surface Laplacian (SL) filter

• Butterworth filter and IIR notch filter ”0.5 to 50 Hz

• bandpass filter,60 Hz notch FIR filter

Signal Decomposition

For decomposition of signal, various types of decomposition method has been used for different author as follows:

• Wavelet Transformation

• DWT (Discrete Wavelet Transformation)

• frequency decomposition

• Fourier analysis (FFT)

• singular value decomposition (SVD)

• linear predictive coding

• TQWT

• Hermite decomposition algorithm

• EMD (Empirical Mode Decomposition

• Adaptive Variational Mode Decomposition

• Clustering technique (CT)

• Empirical mode decomposition (EMD)

• cross-correlation (CC)

• optimal allocated technique

• FAWT

• TQWT

• SPWVD

• ATQWT

• O-VMD

Classification

Key techniques that were used by various author recently are [Vishwakarma et al. 2020]

• AlexNetVGG16 and ResNet50 networks

• ANOVA

• clustering technique-based least square support

• CNN

• cross-correlation based logistic regression algorithm

• DCNN

• Decision-Tree (DT)

• DT

• ELM

• Ensemble Bagged tree

• fuzzy SVM

• HLDA

• kernelized logistic regression (KLR)

• k-Nearest-Neighbour (KNN)

• Least-square support vector machine (LS-SVM)

• leave-one-out cross-validation (LOOCV) with support vector machine (SVM) classifier

• LIBSVM

• Linear-Discriminant analysis (LD)

• logistic regression (LR)

• long-short term memory (LSTM) classifiers

• MCCA

• MC-LS-SVM

• multilayer perceptron neural network (MLP)

• Naïve Bayes

• NDA

• Probabilistic-Neural Network (PNN)

• Random forest

• RCSP for signal decomposition

• RFSVM

• SVM

• SVM RFE

• vector machine

• weighted-kNN

• Z-LDA to LDA

Performance measuring parameter calculation

The performance of the various methods described in this paper, has been assessed using the following factors:

• accuracy,

• Fowlkes-Mallow’s index (FM)

• specificity,

• false alarm rate,

• F1-score,

• Cohen’s Kappa,

• receiver operating characteristic (ROC) curve.

• precision,

• Mathew’s correlation coefficient (MCC),

• sensitivity,

Research gap

Following are the limitation of above methods given by difference authors

• There will be no real time detection.

• No one author work for methodology for low end computing mobile devices.

• No one author work for methodology for low end memory mobile devices.

• Till date no author has automate whole process of detection.

• Till date no author had developed for early prediction of disease/disorders.

• Developed the method using a small number of subjects.

• Techniques utilising empirical hyperparameters for classification

• Parameters for classification and optimisation are chosen at random.

Research challenges

Following are the various challenges now in EEG signal analysis for neurological condition detection:

• About 90% accuracy is achieved.

• In wavelet-based approaches, choosing the appropriate degrees of decomposition and wavelet still leads to hard-band restrictions in filter banks.

• Due to the noise in EEG readings, Higuchi-based algorithms perform poorly.

• The performance of the system must be constrained by a sizeable number of hyperparameters, which must be adjusted according to the CNN techniques. The effectiveness of the approaches described in the literature depends on the parameter. Information is lost and misclassification increases when decomposition parameters are chosen based only on empirical data. To solve these issues, an automated selection of decomposition parameters is needed.

• A fixed width and window function are necessary for the STFT, which also assumes that the signal is steady.

• Wavelet selection is necessary for CWT and DWT, but WVD suffers from cross-terminal frequency and low precision.Solutions from STFT,DWT, CWT and WVD are likewise weakened and noisy as a result of the time and frequency combination.This is resolve by utilising SPWVD by converting the time-series signal to TFR.

• Selecting a cluster size empirically.

• Parameters for optimisation are chosen at random.

Future direction

The future directions outlined in the review encompass the need for further exploration of multimodal data analysis to enhance the accuracy of EEG analysis systems, the application of explainable AI techniques to improve the interpretability and transparency of AI models, investigation of the sustained impact of EEG on cognitive function and overall health, exploration of personalised interventions tailored to individuals’ specific needs, and the development of robust AI models capable of handling variations in data and environmental conditions. Addressing these future directions can advance the field of EEG analysis and contribute to the development of more effective, personalised, and reliable interventions.

Conclusion

In this paper we discussed and compare in detail the models used for classification of EEG signal by various author. Some author (Majdalawieh et al. 2003; Rasheed et al. 2021)– (Péron et al. 2012; Gadhoumi et al. 2016; Kuhlmann et al. 2018; Hosseini et al. 2021) has attempted to review neurological condition but they have coved only one disease/condition but in this paper we have covered about six neurological condition along with detail of every steps and data/methods used. A classifier’s job is to automatically divide a given collection of biological signals into several subgroups. Unsupervised and supervised classification methods are the two main categories of classifications. [4]It happens frequently in biomedical practice that different experts disagree on the earlier labelling. In essence, unsupervised classification gathers related signals into one group such that the differences between groups are as great as feasible. In traditional techniques, the signals are seen as samples of several probability models. The classifier determines a model that the signals meet best after estimating the model variables from the data. If the inputs do not fit the assumed probability models, the classifier performs worse.

Declarations

Conflict of Interest On behalf of all authors, the corresponding author states that there is no conflict of interest.

Disclosure statement

No potential conflict of interest was reported by the authors

Data availability statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study

Additional information

Notes on contributors

Abhishek Jain

Mr. Abhishek Jain is currently working as an Assistant Professor in Information Technology Department, Central University, Bilaspur, India and having 14 years of industrial, teaching and research experience. He has completed B.E. from GEC Raipur and M. Tech from NIT Allahabad currently pursuing PhD from Central University Bilaspur C.G. He is also worked as senior software engineer in TCS. He has authored more than 10 scientific publications in peer reviewed journals. He has worked in Geographical Information Systems (GIS) project Thames Water London UK and RAPDRP Gujarat. His industrial working experience include object-oriented analysis modelling and design, database administrator, Oracle, SQL, XML, XSLT, DTD, Java etc. His research focused on remote sensing, image processing, signal processing biomedical engineering, EEG signal, ECG Signal, IOT, block chain technology.

Rohit Raja

Dr. Rohit Raja is working as Associate Professor & Head in the IT Department at the Guru Ghasidas, Vishwavidyalaya, Bilaspur (CG) has done Ph.D. in Computer Science and Engineering in 2016 from C. V. Raman University India. His main research interest includes face recognition, Signal Processing, and Networking and Data Mining. He has filed successfully 27 patents. He has been received 3 times invitation being a Guest in Scopus Indexed IEEE/Springer Conferences. He has been invited many times being an expert in various Colleges/universities in India. He has published 100 research papers in various International/National Journals (including IEEE, Springer, etc.) and Proceedings of the reputed International/ National Conferences (including Springer and IEEE). He has received the “Excellence in Research” Award under the Category of HEI Professor by Auropath Global Awards, 2019. He has received the “Distinguished Research Award “I2OR Awards 2019 during the 4th International Conclave on Interdisciplinary Research for Sustainable Development 2020 (IRSD 2020). He has received 2nd Runner up an award in GRAND Finale of SMART INDIA HACKATHON 2017 Ministry of Skill Development and Entrepreneurship, Government of India. He is a member of a professional international society including IEEE, ISTE, CSI, ACM, etc. He is on the editor/ reviewers of many peer-reviewed and refereed journals (Including IEEE, Springer, ender science).

Sumit Srivastava

Dr. Sumit Srivastava is working as an Assistant Professor in the department of Computer Science and Engineering, Birla Institute of Technology Mesra, Ranchi, India. He has done his Ph.D. in Computer Science and Engineering from the Birla Institute of Technology Mesra, Ranchi. He has 12 years of experience in teaching and research. His areas of interest are speech processing, signal processing, cryptography and network security, embedded Systems, and machine learning. He has received several patents. He has also published several papers in various international/national journals and the proceedings of prestigious international/national conferences.

Prakash Chandra Sharma

Prakash Chandra Sharma is currently working as Associate Professor in the School of of Information Technology at Manipal University Jaipur, India. He has received his honored degree of Ph.D. from Indian Institute of Technology Indore in discipline of Computer Science & Engineering. He completed his Master of Engineering in Computer Engineering discipline from SGSITS Indore. He has experience of teaching and research of more than 18 years in various areas of Computer Science & Engineering. His research interest includes Graph Theory, Theoretical Computer Science, Artificial Intelligence, and Soft Computing. The key subjects, he has taught at UG and PG levels include Data Structures & Algorithms, Theory of Computation, Operating System, Compiler Design, Artificial Intelligence, Semantic Web, Graph Theory, Soft Computing, Discrete Mathematics and many more. He has published number of research papers in various International Journals indexed in SCI/Scopus, and in Proceedings of the reputed International/ National Conferences as well. He has also published number of book chapters in various edited book with reputed publishers like Elsevier, Springer, CRC Press, Willey etc. He is active reviewer and editorial member of various reputed International Journals in his research areas. He is a member of various Professional International Societies (including IEEE, CSI, IAENG, Institute of Engineers India). He is recipient of many awards and recognitions at national level.

Jayesh Gangrade

Jayesh Gangrade received PhD degree in Computer Science & Engineering from Maulana Azad National Institute of Technology Bhopal in 2020. He has received MTech. degree with honours in Computer Science and Engineering from Maulana Azad National Institute of Technology Bhopal in 2008. He has also worked as a project engineer in research centre in Wipro technology Bangalore. He has more than 18 years teaching and research experience in various areas of Computer Science & Engineering. He is currently working as an Associate Professor at Manipal University Jaipur, Rajasthan. He has published several research papers in various International/National Journals and in Proceedings of the reputed International/ National Conferences as well. His research interests are in the areas of computer vision, image processing and Deep learning

Manoj R

Manoj R, an academician with B.E., MTech and currently pursuing his PhD in CSE in block chain technology. He is working as an assistant professor at the Department of Computer Science & Engineering at Manipal Institute of Technology, Manipal Academy of Higher Education (MAHE), Manipal, Karnataka, India. He has an academic experience of more than 15 years. He has presented and published more than 20 papers in reputed journals and conferences of international repute. He has reviewed many research papers for reputed journals and conferences. His research interests cover cryptography and network security, and block chain and its applications.