1. Introduction

The brain computer interface (BCI) is one of the currently evolving technologies that have attracted significant attention over recent decades and have witnessed remarkable improvement both in speed and accuracy [1]. By definition, BCI is simply a hardware and software communications system that enables humans to interact with their surroundings by directly acquiring and analyzing neural signals between the brain and the computer. Unlike the conventional systems which are controlled by computer, the BCI is controlled by human brain signal [2]. Basically, BCIs are of active types which are controlled by means of endogenous tasks such as motor imagery and mental arithmetic operations, and reactive types that are controlled using external stimulation like auditory, visual and haptic [3].

BCI have been reported to have facilitated restoration of the movement ability for physically challenged or locked-in users and replacing lost motor functionality. At present, BCI have been proposed as a tool for diagnosing, treating and following up many other neurophysiological and neuropsychological disorders [4]. The recent research trends have applied BCI to non-medical applications for instance, normal subjects explores BCIs as a novel input device and investigation of the generation of hands-free applications. Also, BCI permits re-integration of the sensory–motor loop [5].

Irrespective of its type, a BCI is basically made up of the signal acquisition module, translation algorithm module, control interface module and device controller module. Of all the modules, the translation algorithm is an important tool for detecting brain activities and abnormalities thus; the current research focuses on the problem of EEG signal pattern, control signal transfer algorithm and system application [7]. The control applications of BCI have been an object of intensive research by researchers resulting in improvement in product developments [8, 10]. For instance, in 1999, a patient could type 0.5 characters per minute through slow cortical potential (SCP) BCI [6] while in 2007, a commercial speller controlled by visual attention averaged 7.5 characters per minute [13].

EEG is one of the mostly used non-invasive modalities for probing the human brain functions however; none of these modalities is effective in providing necessary information to understand the spatio-temporal aspects of information processing in the human brain [19, 9]. While [21, 20] defined EEG as a representation of post-synaptic potentials that are generated at cortical level by synchronous activity of about 105 (10 rates to 5) neurons or the electrical activity recorded from the human scalp [9]. EEG is widely used in BCI for detection and analysis of abnormalities of the brain [23].

Though EEG is excellent in terms of its inherently high temporal resolution and precision, it’s instantaneous, nonlinear and non-stationary nature results in low spatial resolution and specificity [18]. Attempts to solve the problem of low spatial resolution in EEG includes [24, 26] methods to improve the spatial resolution of EEG likewise [25] spatial filter approach for evaluation of the surface Laplacian of the EEG and MEG. Also, [27] employed the lead field theory approach to address the problem of spatial resolution in EEG. Apart from the fact that EEG contains artifacts produced by eye movements and/or blinks [20, 29] it also has redundant or noise information both from the subject and equipment interferences [30, 21].

The feature extraction is a critical issue in translation algorithm development as it considers the extraction of the most discriminative features so as to arrive at a system with high performance. Above all, BCI still faces a lot challenges that results in performance variation across and even within subjects [13]. Thus, this work provides a concise but all compassing review of methods that have been adopted in the recent time for development of an EEG classification in BCI.

2. Electroencephalography (EEG)

The electrical activity of neurons is response for the formation of EEG. The first human EEG signal was recorded by Hans Berger in 1924 by placing two electrodes on a patient’s skull to detect a very weak current using a galvanometer. The EEG is a noninvasive procedure for registering the brain activity through digital recordings thus, EEG has provided promising ways for computer-based signal processing to aid in epilepsy diagnosis [20, 9]. The EEG signal is the resulting waveform representing the overall electrical activity of the brain arising from many neuronal activities. Nowadays, clinical EEG machines can be found in many clinics for routine brain electrical activity monitoring and assisting the physicians in decision-making processes [16]. The EEG has emerged as a fundamental tool in the diagnosis and research of several brain disorders, including those related to epilepsy. A clear understanding of the basics of EEG signal generation and recording is necessary as in Figure 1 in order to effectively model and analyze EEG.

Figure 1. EEG Signals of Emanating from Lobes of the Brain (Source: Richard and Michael, 1974)

The EEG signal from the scalp is typically characterized by amplitude of approximately 100 μV and time duration of 0.01–2 s [32, 33]. EEG signals are basically affected by moods such as drowsiness, excitement, and relaxation. In clinical tests, recording electrodes are placed on the scalp using international standards EEG geometrical sites, such as the 10–20 system [28]. Clean contact between the electrode and the skin is necessary for good EEG recording; the conductive gel is usually applied to reduce the impedance between the electrode and the scalp. Each electrode is input to a differential amplifier commonly set between 1,000 and 100,000 amplification.

The temporal resolution of EEG is about 1msec which mean that events of short duration, such as epileptic spikes that which lasts for about 1msec can be reliably recorded. Although normal EEG fluctuations have amplitudes of 75µV or more, the magnitude of useful brain signals buried in these fluctuations is often considerably smaller [20, 9]. The electrical activity of the brain recorded in an EEG is normally distributed in a few frequency ranges, corresponding to different brain states. Rhythmic sinusoidal activities can be recognized within the EEG signal; the frequency compositions of the EEG signals commonly used for analysis and are categorized into main five frequency bands as shown in Table 1.

Table 1. Brain Rythms and Respective Frequency Band

EEG has been widely used by neuroscientists to study brain function. The mathematical modeling and analysis of EEG has advanced the development of computer interface tools that assisted the identification of salient patterns embedded within the EEG to improve recognition. Aside from the fact that BCI can facilitate the communication of physically handicapped individuals with the help of a computer by using EEG signal characteristics, BCI has been recently employed in the restoration of the disable people’s day-to-day activities and in the diagnosis and monitoring of Alzheimer disease (AD), Epilepsy, Huntington’s disease (HD) and Sleep disorders. Other applications of EEG are in evoked potentials or evoked responses that are useful for evaluating a number of neurological conditions.

3. Review of Methods

This section presents review of some of the methods that have been employed by researchers in BCI and EGG classification researches in the recent times.

3.1. Principal Component Analysis (PCA)

PCA is a useful statistical technique that is based on linear transformations which maps data from high dimensional space to low dimensional space and it has found application in fields such as image recognition and compression [65, 66]. PCA is a powerful method in image formation and data patterns used for feature extraction in which similarities and differences between them are identified efficiently [67, 68]. It has the advantage of reduced dimension by avoiding redundant information without much loss [69]. It achieves this by defining new co-ordinate axes in directions that are rich in information content. The transform variables with the largest value can be assumed to have the greatest discriminatory power [75]. The Eigen vectors of the covariance matrix determine low dimensional space. If x(0), x(1), ….. x(n-1) is a set of input sample and x is the N x 1 corresponding vector

(1)

Also

(2)

Where H = denotes Hermittan operation, y = is the transformed vector of x and A = unitary matrix N x N.

From the definition of unitary matrix the above equation become

(3)

A, ai, i = 0, 1,…N-1 columns are called the basis vector of the transform. By PCA definition;

(4)

Where λ_i are the eigenvalues and is the largest eigenvalue of the correlation matrix.

(5)

This equation 5 generates features that are mutually uncorrelated.

3.2. Independent Component Analysis (ICA)

Independent component analysis (ICA) is a computation method for separating multi-source signals into subcomponents, with the assumption that the signals are mutually statistically independent [70]. This analysis is a type of blind source separation that determines the independent components by maximizing the statistical independence of the estimated components [50]. The ICA theory goes beyond PCA in that it try to achieve much more than simple decorrelation of the data. Given the input samples x, the invertible matrix W of N x N dimensions with the following observations y(i), i = 0, 1, . . . , N - 1, of the transformed vector are mutually independent given as

(6)

The goal of statistical independence is a stronger condition than the uncorrelatedness required by PCA. ICA has the capability of revealing information from the higher order statistics of the data [71]. Assuming that the input random data vector x is indeed generated by a linear combination of statistically independent and stationary components such that

(7)

then A is the mixing matrix and W is the de-mixing matrix.

3.3. Fast Fourier Transform

The development of the fast Fourier transform (FFT) led to extreme popularity of the Fourier transform. The FFT reduces the computation effort from N² for the conventional discrete Fourier transform to N log₂N in one dimension for the FFT which translates to efficiency gain. For instance, assume that a two-dimensional fast Fourier transform of a 1024 × 1024 pixel image takes 5 seconds on a computer; the conventional discrete Fourier transform would take 14 hours for the same image. The operation of FFT can be represented as in the following equations

(8)

Equation 8 can be re-written as in 9

(9)

According to [35], the summation can be split into odd and even parts as in:

(10)

This equation reveal that the N-point DFT reduces to two N/2-point DFTs. The symmetry of the phase shifts comes into play when the number of summations is reduced from N to N/2.

3.4. Self-Organizing Maps (SOM)

Self-organizing feature maps is an unsupervised neural network and is closely related to multidimensional scaling. The method is very useful when there is a nonlinear mapping inherent in the problem itself. SOM provides mechanisms for visualizing the complex distribution of cognitive states. High-dimensional data are projected to low-dimensional data using two layers: the input layer and the output layer [50]. It represents all points in the source space by points in a target space such that distance and proximity relationships are preserved [72]. The learning process of the SOM is similar to the information representation properties of many functions of the brain. The modes in the target space computes its net activation by

(11)

If the most activated unit is denoted by y*, the weight to this unit and its immediate neighbourhood are updated by

(12)

Where n(t) is the learning rate that depends on t, t is the iteration number, is the window function which ensures that neigbouring points in the target have weights that are similar and its value is 1.0 for y = y*.

3.5. Wavelets

Wavelet transform is a multi-resolution technique that offers the advantage of time-frequency representation of the image [65]. Its approach is such that it breaks up a signal (image) into shifted and scaled versions of the “mother” wavelet. Wavelet analysis is done by convolving the signal wavelet kernels to obtain wavelet coefficients representing the contributions of wavelets in the signal at different scales and orientation [84]. The wavelet transform is very popular since it allows for localization in time and frequency [74, 80, 85]. These basis functions include Haar (haar), Daubchies (db), Symlets (sym), Coiflets (coif), and biorthogonal (bior). They have compact support but differ in properties thus, the selection of the basis function is a key issue in a wavelet transform based analysis. Consider a signal x(t), its discrete wavelet transform DWT can be represented as follows:

(13)

Where are the scaling coefficients, is a scaling function, are the wavelet coefficients and is the wavelet function. The scaling and wavelet function can further be express a family of functions as follows

(14)

(15)

The basis function listed above can be used to decompose data into different resolutions from which approximation and detail coefficients can be computed.

3.6. Distance Classifier

The distance classifier is a method of classification in which the classes are similar in distribution and are linearly separable with the underlying observation model for each sample described by a reflectivity. The approximated probability density function (PDFs) generated using this reflectivity parameter can be compared with the other PDFs in the database. The decision rule according to [81];

(16)

Where w₁ is the class 1, w₂ is the class 2, X is the sample, M₁ is the mean of class 1, M₂ is the mean of the class 2, P₁ is the probability of the class 1 and P₁ is the probability of the class. Hence the decision lines are allocated half way between the centers of clusters of different classes i.e. P₁ = P₂ = 0.5. The distance classifier can be computed using the following algorithm:

1. Group the data set into supervised number of classes to be considered clusters according to their labels.

2. Estimate the sample means for each class by averaging the parameter set of the class.

3. Classify test sample by assigning it to the class which has the nearest means vector.

4. Estimate error rate by the percentage of misclassified samples.

3.7. k-Nearest Neighbor (kNN) Classifier

The k-Nearest neighbors (kNN) classifier is a nonparametric method that classifies a test sample to the class of the majority of its k-neighbors [65, 81]. It is a nonlinear classifier and assuming the number of voting neigbours to be k= k₁ + k₂ …k_N, where k_i is the number of samples from i in the k– sample neighborhood of the test samples. The test sample is assigned to class e if

(17)

Based on the following algorithm voting kNN classifier can be computed as follows:

1. Get the distances between the test sample and the samples in the design set and store it.

2. Arrange the obtained distances values in ascending order.

3. Take the subset of the first k distance in the sorted array; i.e. kNN.

4. Estimate error rate by comparing the classification result with actual class membership.

The computation complexity of the kNN algorithm both in space and time has received a great deal of attention. However, the greatest use of kNN techniques is for problems with many features thus, attention is given to the general d-dimensional case [12]. The prestructuring, computing partial distances and editing the stored prototypes are the general algorithmic techniques for reducing the computation burden in k-nearest-neighbor searches.

3.8. Support Vector Machine (SVM) Classifier

SVM clustering is a state-of-the-art learning machine that utilizes statistical learning theory [83], which was originally proposed for classification of data. It aims at defining an optimal hyperplane, that separates the training data so that the minimum expected risk is achieved [73]. Unlike the conventional neural networks the SVM is flexible in that it has many parameters that can be adjusted to achieve better classification rate [77, 78]. Different from other classifiers, SVM is less affected by the so-called “curse of dimensionality” [79]. According to [81], linearly separable classes can be described by a hyperplane:

(18)

This implies that the support vectors lie on either of the two hyperplanes and they form the critical elements of the training set.

3.9. The Neural Networks

The artificial neural network is capable of receiving stimulus from other neurons and can send a reaction to a number of neurons [76, 82]. It has a powerful capability to create hyperbolic surfaces in addition to the original straight lines as decision boundaries [86]. They have been proven themselves to be proficient classifiers and are well suited for tumour classification [87, 88]. Based on [86], the bipolar neural activation function is taken to be of the form

(19)

Where net = sum of all inputs neuron multiplied by their weight, λ = activation constant usually one for all architecture. Back propagated neural networks is one of the most well known and oldest learning techniques in which back propagation algorithm is used for training the neural network [89, 11]. The advantage of the back propagation algorithm is that it has a number of parameters that can be varied in order to optimize performance of the classifier. These parameters include the momentum and the learning rate.

4. Brain Computer Interface (BCI)

Brain computer interface (BCI) is one of the currently evolving technologies that have attracted significant attention over recent decades and have witnessed remarkable improvement both in speed and accuracy [1]. BCI is simply a hardware and software communications system that enables humans to interact with their surroundings by directly acquiring and analyzing neural signals between the brain and the computer. BCIs are basically devices that translate changes of the neurophysiological activity of the brain into control commands for an application [15]. Unlike the conventional systems which are controlled by computer, the BCI is controlled by human brain signal [2]. The central element of a BCI is the translation algorithm that converts electrophysiological input from the user into output that controls external devices.

Basically, BCIs are of active types which are controlled by means of endogenous tasks such as motor imagery and mental arithmetic operations, and reactive types that are controlled using external stimulation like auditory, visual and haptic [3]. According to [35], a BCI can generally be divided into three classes namely; the sensory interfaces, the cognitive interfaces and motor interfaces.

The communication channel of BCI is potentially useful in emerging researches that ranges from psychology and computational neuroscience to engineering such as bioengineering, human subject monitoring, neuroscience research, man – machine interaction and so on [1, 23]. For instance, BCI creates a new non-muscular channel for relaying a person’s intentions to external devices such as computers, speech synthesizers, assistive appliances, and neural prostheses. Also, BCI permits to re-integration of the sensory–motor loop [5]. At present, BCI have been proposed as a tool for diagnosing, treating and following up many other neurophysiological and neuropsychological disorders [4]. BCI represents a new frontier as an interdisciplinary research direction [7, 36] since it works with neuroimaging techniques that plays critical role in neuroscience research and management of neurological and mental disorders [34].

In [51], a BCI based on electrocorticographic (ECoG) was worked upon to enable users control a one-dimensional computer cursor rapidly and accurately and finally suggested that an ECoG-based BCI could provide a non-muscular communication and control for subjects with severe motor disabilities. Meanwhile, [57] worked on common spatial pattern (CSP) to decorrelate EEG signals obtained from different electrodes and classification results indicate WCSP outperforms CSP for the true asynchronous BCI system with an average Kappa increase of 0.4. In the work of [52], BCI System bit rate for controlling a virtual telephone keypad was developed by a simulated virtual telephone keypad based on Steady State Visual Evoked Potential (SSVEP) using dynamic programming technique with a conclusion that that user reached requirement faster with little number of selections and thus increased transfer rate. In [56], a non-invasive BCI for the decoding of intended arm reaching movement in prosthetic limb control was designed and create a signal decoding strategy that allows more command over potential prosthetic devices thereby improving the classification accuracy from 60.11% to 93.91% in the binary class.

In [54], a work on the control of BCIs by users with cerebral palsy was conducted with 14 individuals with CP attempting to control two standard online BCIs based upon sensorimotor rhythm modulations and based upon steady state visual evoked potentials. According to [55], a novel hybrid BCI system that combines motor imagery (MI)-based bio-signals and steady-state visual evoked potentials (SSVEPs) to control the speed and direction of a real wheelchair synchronously was poposed and the results validated the efficiency of the developed system with an accuracy rate of more than 85% for all subjects.

In [47], training leads to increased auditory BCI performance of end-users with motor impairments in which a newly auditory BCI paradigm with natural sounds and directional cues was developed. The two best end-users achieved information transfer rates of 5.78 bits/ min and accuracies of 92%. In [58], a modified version of the CC-LR algorithm that explore a suitable feature set was developed and it was reported that the proposed method outperforms the recently reported eight methods for the MI tasks EEG signal classification. While [59], presented a hybrid mental speller that can effectively prevent unexpected typing errors based on the steady-state visual evoked potential (SSVEP) and the result of the online experiments showed that the system could significantly reduce the total typing time thus, enhancing the performance of the speller by preventing typing errors.

A telematics and informatics BCI system that consists of discriminative area selection, feature extraction and classification was proposed by [60] and the average classification accuracies of the three datasets used are 85.6, 83.1, and 81.3%, respectively. According to [63], a novel hybrid BCI system that uses near infrared spectroscopy (NIRS) and EEG was presented. The results of an online experiment demonstrated that the proposed system had a true positive rate of about 88%, a false positive rate of 7% with an average response time of 10.36 s. In [63], a prototype to test real time data collection and navigation through interface by detection and classification of event-related potentials (ERPs) was presented. The time segment (TS) in combination with LDA produces the best results for all subjects giving an average of 85% accuracy.

In the work of [61], a shift invariant ERP detection strategies on data from ten subjects obtained in a P300 speller experiment was proposed. The results support the conclusion that ERP detection can be achieved without a precise knowledge of the stimulus onsets. For [45], an online three-class transcranial doppler ultrasound BCI in which vision-independent right-lateralized tasks were investigated. It was concluded that the results demonstrated the potential of a three-class online TCD BCI that does not require visual task.

Table 2 presents a summary of the review of related works on EEG based BCI. The works presented in the Table were between 2004 and 2016; with 31.25% of the works done in 2015. It was also reflected from the work that 37.5% of the works in the table focused on control. The highest classification accuracy of 93.91% was reported in the table while 74% is the lowest reported.

Table 2. Summary of Review of Related Works on BCI

4.1. Feature Extraction

The accuracy or efficiency of a classification system depends largely on the feature(s) of the samples to be classified that is supplied to it [12] hence; feature extraction is an important stage in BCI. Feature extraction is the process of deriving new features from the original features in order to reduce the cost of feature measurement, increase classifier efficiency and allow higher classification accuracy [14, 37]. According to Deserno 2011, there are different levels of feature extraction namely; data level, pixel level, edge level, texture level [11] and region level. In attempt to correctly extract features for pattern recognition system, a number of transforms have been applied successfully. It has been discovered that different transform suits different application. However, most medical CAD systems are based on texture features which are extracted from biomedical images [38]. Most of the commonly used transforms for feature extraction include Wavelets, FFT, LDA, PCA, EMD and SOM.

As presented in Table 2, [53] worked on a feature selection algorithm to detect subject-dependent feature and channel relevance for mental task discrimination. In [48], the potential of Sample Entropy (SampEn) as a feature extraction method for automatic epilepsy detection and classification of normal, interictal, ictal and epileptic seizures EEG signals was investigated with a reported classification accuracy of 95.67%. In [44], a BCI with genetic algorithm was developed based on entropy and wavelet transform for feature extraction and two neural networks, including probabilistic neural network (PNN), Multilayered Perceptron (MLP) and support vector machine (SVM) were employed and their results were compared. A work on automatic EEG seizure detection using dual-tree complex wavelet-Fourier features was reported by [46] to achieve perfect classification rates (100%) for the EEG database from the University of Bonn. It was therefore concluded from the work that the conventional FFT could be replaced by sparse FFT so that the proposed method could be even faster. In [42], an epileptic seizure detection system that analyzes EEG signals using different transformation techniques and decompositions was presented to achieve an average sensitivity of 91.36%. In the same manner [43], presented epileptic seizure prediction system based relative spectral power features to improve sensitivity and specificity of prediction methods. The best results demonstrated a sensitivity of 75.8% and a false prediction rate of 0.1per hour.

4.2. Classification of EEG Signal

The central element in each BCI is the classification module which is also referred to as translation algorithm. It simply converts electrophysiological input from the user into output that controls external devices. The translation algorithm is an important stage in the signal processing module of the BCI system and it is responsible for translating the extracted signal features into device commands that performs the user’s intent. Whatever the nature is, a translation algorithm changes signal features into device control commands. The first part of signal processing simply extracts specific signal features. The extracted signal features may be classified on both frequency and shape features based on linear methods or nonlinear methods like the neural networks [17].

BCI operations have been said to depend mainly on effective interaction between two adaptive controllers the user who encodes his or her commands in the electrophysiological input provided to the BCI, and the computer which recognizes the command [7]. The development of translation algorithms solely relies on the classifiers like kNN, LDA, Neural Network and SVM.

In [39], a neural network classification of EEG signals using AR with MLE preprocessing for epileptic seizure detection was presented with better specificity of 96.2% of the patients as epileptic seizure patients. The development of an EEG preprocessing technique that significantly improved the sensitivity and specificity of EEG based of detection of Alzheimer's disease (AD) was addressed in [40]. While [41] proposed a multiclass support vector machines (SVM) for EEG-signals classification with the error-correcting output codes to achieve 93.630%, [30] worked on detection of brain tumor in EEG signals with SVM as classifier that enables effective and early detection and classification of brain tumors thus initiating quicker clinical responses.

The work of [49] simply aimed at distinguishing the normal and abnormal hearing subjects using acoustically stimulated EEG signals to achieve a classification accuracy of 96.75% was obtained. In [64], a drowsiness detection mechanism was developed based on an EEG collected from the driver with an off-the-shelf mobile sensor and a success rate close to 80% was achieved based on the EEG data. Artifacts removal and selection of useful brain sources were done based on the ICA drivers EEG while a SOM was employed to recognize all distracted and concentrated EEG epochs in [50]. It was concluded from the work that the proposed BCI system reached a maximum accuracy of approximately 90% for the recognition of EEG epochs. The work of [48] was centered on the development of an automated classification of EEG signals for the detection of epileptic with overall classification accuracy of 99%.

Table 3 presents a summary of the review of related works on EEG based neural abnormalities detection, prediction and classification. The works reviewed in the Table were published between year 2005 and 2015 with 41.67% of the works published in 2014. The table also reflected that 83.33% of the works focused on detection and classification. Also, 50% of the works used SVM for classifying the EEG signals. FFT also took 50% of the feature extraction techniques. The highest classification accuracy of 100% was reported in the table while 63.6% is the lowest reported.

Table 3. Summary of Review of Related Works on EEG Based Neural Abnormalities Detection, Prediction and Classification

Though the existing works reviewed in this paper achieved good classification accuracies, none of the works considered the use of hybridized classifier. Future researches should consider the use hybridized classifier for classifying EEG data.

5. Conclusions

This work is successful in providing a concise but all compassing review of methods that have been adopted in the recent time for development of an EEG classification in BCI. The work established that SVM and FFT are the mostly adopted classifier and feature extraction method for the development of EEG classification system. It is evident from this review of literature that there are rooms for improvement which may be achievable through classifier hybridization and improved feature extraction methods.