Discrimination of Stroke-Related Mild Cognitive Impairment and Vascular Dementia Using EEG Signal Analysis

Noor Kamal Al-Qazzaz 1,2*, Sawal Hamid Bin Mohd Ali 1, Siti Anom Ahmad 3,4, Mohd Shabiul Islam 5and Javier Escudero 6

1 Department of Electrical, Electronic & Systems Engineering, Faculty of Engineering & Built Environment, Universiti Kebangsaan Malaysia, UKM Bangi, Selangor 43600, Malaysia; sawal@ukm.edu.my

2 Department of Biomedical Engineering, Al-Khwarizmi College of Engineering, Baghdad University, Baghdad 47146, Iraq

3 Department of Electrical and Electronic Engineering, Faculty of Engineering, Universiti Putra Malaysia, UPM Serdang, Selangor 43400, Malaysia; sanom@upm.edu.my

4 Malaysian Research Institute of Ageing (MyAgeing), Universiti Putra Malaysia, Serdang, Selangor 43400, Malaysia

5 Faculty of Engineering, Multimedia Universiti, MMU Cyberjaya, Selangor 63100, Malaysia; shabiul.islam@mmu.edu.my

6 Institute for Digital Communications, School of Engineering, The University of Edin-burgh, Edinburgh EH9 3FB; UK; javier.escudero@ed.ac.uk

* Author to whom correspondence should be addressed; noorbmemsc@gmail.com; Tel.: +964-773-543-1383.

Abstract: Stroke survivors are more prone to developing cognitive impairment and dementia. Dementia detection is a challenge for supporting personalized healthcare. This study analyzes the Electroencephalogram (EEG) background activity of 5 vascular dementia (VaD) patients, 15 stroke-related patients with mild cognitive impairment (MCI), and 15 control healthy sub-jects during a working memory (WM) task. The objective of this study is twofold. First, it aims to enhance the discrimination of VaD, stroke-related MCI patients and control subjects using fuzzy neighborhood preserving analysis with QR-decomposition (FNPAQR); second, it aims to extract and investigate the spectral features that characterize the post-stroke dementia patients compared to the control subjects. Nineteen channels were recorded and analyzed us-ing the independent component analysis and wavelet analysis (ICA−¿WT) denoising tech-nique. Using ANOVA, linear spectral power including relative powers (RP) and power ratio were calculated to test whether the EEG dominant frequencies were slowed down in VaD and stroke-related MCI patients. Non-linear features including permutation entropy (PerEn) and Fractal dimension (FD) were used to test the degree of irregularity and complexity, which was significantly lower in patients with VaD and stroke-related MCI than that in control subjects (ANOVA, p ˂ 0.05). This study is the first to use fuzzy neighborhood preserving analysis with QR-decomposition (FNPAQR) dimensionality reduction technique with EEG back-ground activity of dementia patients. The impairment of post-stroke patients was detected us-ing support vector machine (SVM) and k-nearest neighbors (kNN) classifiers. A comparative study has been performed to check the effectiveness of using FNPAQR dimensionality reduc-tion technique with the SVM and kNN classifiers. FNPAQR with SVM and kNN obtained 91.48% and 89.63% accuracy respectively, whereas without using the FNPAQR exhibited 70% and 67.78% accuracy for SVM and kNN respectively, in classifying VaD, stroke-related MCI, and control patients, respectively. Therefore, EEG could be a reliable index for inspect-ing concise markers that are sensitive to VaD and stroke-related MCI patients compared to control healthy subjects.

Keywords: electroencephalography; ICA−¿WT; relative power; permutation entropy; fractal dimension; vascular dementia, mild cognitive impairment

1. IntroductionStroke is a cerebrovascular disease considered the leading risk factor for severe physi-

cal disability and cognitive impairment [1]. Approximately 30% of stroke patients develop dementia within the first year of stroke diagnosis [2]. Vascular dementia (VaD) is the second most common case of dementia after Alzheimer's disease (AD); between 1% and 4% of el-derly people aged 65 years suffer from VaD, whose prevalence doubles every 5 to 10 years after this age [3]. Mild cognitive impairment (MCI) is clinically defined as a decline in cogni-tive function (mostly memory) that is more significant than expected with respect to the indi-vidual’s age and education level, but the patient still lives independently [4]. MCI is consid-ered as a stage between early normal brain cognition and late severe dementia [5].

Working memory (WM) is severely affected in this case and it is defined as the ability to provide a temporary storage and to manipulate information for complex cognitive tasks such as attention, comprehension, reasoning, planning, and learning within a short period (10–15 seconds up to 60 seconds) [6]. The WM capacity of an individual can store temporar-ily approximately 7 ± 2 items [6, 7].

Electroencephalogram (EEG) is a non-invasive technique that used to record cortical brain activity. EEG can be classified into five frequency bands: delta (δ), theta (θ), alpha (α), beta (β), and gamma waves (γ) [5]. However, EEG is affected by non-cerebral sources that can mimic brain cognitive or pathological activity and overlap with EEG frequencies [5]. Therefore, the EEG data are exposed to contamination with various artifacts.

The artifact’s amplitude is relatively large with respect to the cortical signal. The arti-facts can be divided into physiological artifacts that originate from the generator sources within the body (e.g., heart, eye, and/or muscles) and non-physiological artifacts (technical origin) that are related to the environment or equipment [8, 9]. The physiological artifacts can be muscle artifact (MA), cardiac artifacts (CA) and ocular artifacts (OA). The non-physiolog-ical artifacts [9] include the power line interference and sweat artifacts.

Many researchers have used independent component analysis (ICA) to separate dis-tinct artifacts from EEG signals efficiently. For instance, Sameni et al. [10] applied the ICA algorithm to remove electrooculogram (EOG) artifacts from multi-channel EEG recordings. Meanwhile, wavelet transform (WT) is a common and powerful denoising method that is widely applied to biomedical signals. WT has also been extensively utilized because it can re-move electromyogram and EOG noise. Discrete WT (DWT) has also been considered a promising technique to represent EEG signal characteristics by extracting features from the sub-band of EEG signals [11].

Recently, using a combination of denoising methods from EEG has gained attention for multi-channel processing [12]. For instance, Mammone et al. [13] used DWT to partition each channel of the original EEG into the four bands of human brain activities. Each band of each channel was represented by WT components, and the WT components linked to artifacts were automatically identified and cascaded into ICA. Artifactual components were rejected before conducting inverse ICA (inv–ICA) and inverse DWT (IDWT) [13]. An ICA–WT tech-nique has been used in this study to denoise critically marked ICs using DWT and to recon-struct ICA-corrected EEG signals. Therefore, the advantage of the proposed technique is to detect and extract artifactual components using a completely automatic EEG artifact detector based on the joint use of spatial and temporal features (ADJUST) algorithm [14] that identi-fies the artifactual ICs on the scalp topographic maps obtained by EEGLAB version 13.0.1b [15].

In order to develop a non-invasive useful diagnostic index which would be valuable for early dementia diagnosis and to discriminate dementia degree of severity, it is necessary to enhance the system interpretability specifications that become helpful in improving the de-mentia detection.

In this setting, spectral analysis has intensively been used to detect abnormalities in the spectra of dementias’ patients EEG. To quantify these changes, the slowing activity of EEG signal in dementia can be summarized by a shift in power spectrum toward low fre-quency, namely increase of power in low frequency (i.e δ and θ bands) and a decrease of power in high frequency (i.e α and β bands) [16]. Further to this, Moretti et al. have proposed several EEG markers that could have a prospective value in differential diagnosis between vascular and degenerative MCI. For example, the θ /α ratio could be as a marker for the esti-mation of the individual extent of cerebrovascular damage related to dementia. Additionally, the increase of θ /γ ratio is best associated with amygdalar atrophy [17-20]. Moreover, Schmidt et al. have calculated the α / β ratio of the mean potential of EEG represents a good marker discriminating AD patients from normal controls [21]. Add to that, Lundqvist et al. have summarized the effect associated with growing memory load In increasing θ and γ power and decreasing α / β [22].

As the brain can demonstrate a non-linear behavior, this study also employs non-lin-ear analysis with conventional linear methods to provide additional information about mental diseases [23, 24]. Entropy is a powerful concept for evaluating non-linear dynamic character-istics of a signal; the uncertainty and irregularity of a time series are measured using high en-tropy, which generally demonstrates the high uncertainty of the system, or low entropy, which shows the high regularity and certainty of the system [25, 26]. Permutation entropy (PerEn) can also be used for both non-stationary and non-linear signals [26, 27]. Morabito et al. [28] used PerEn to assess the irregularity of EEG recordings in Alzheimer’s disease.

Fractal dimension (FD) is a complexity metric that can be obtained by evaluating scale-free (fractal) properties. FD is used to quantify the complexity and self-similarity of MEG and EEG time series [29]. Therefore, under the hypothesis that the FD of the EEG is sensitive to the neuronal dysfunction associated with a brain lesion, this study tests the ability of Higuchi’s fractal dimension (HFD) to assess patients with VaD and stroke-related MCI.

Pearson's correlation of resting-state EEG signals of different scalp regions may be re-duced in MCI and AD patients, compared to control subjects [30].

Many research efforts have done in the stages of feature extraction, dimensionality re-duction and classification for achieving valuable discriminations. Therefore, feature vectors must be analyzed further before being applied to the classifier to avoid overloading it, reduce the computational time, and increase the classification accuracy. Feature dimensionality reduc-tion is widely used method to avoid the potential redundancy of high-dimensional data [31-33]. The dimensionality-reduced features were used as an input to the classifiers to improve the accuracy of the classification of the severity of dementia by EEG signal analysis. Linear discriminant analysis (LDA) and fisher discriminant analysis (FDL) have been widely used for their fast and simple implementation, suitable for real-time [34]. Their objective is to create a new variable that combines the original predictors by finding a hyper-plane that separates the data points representing different classes and that minimizes the variance within the class un-der the assumption of normal data distribution [31]. This study employed the fuzzy neighbor-hood preserving analysis with QR-decomposition (FNPAQR) dimensionality reduction tech-nique [35] of Khushaba et al. [35] to minimize the distance between the samples belonging to the same class while maximizing the distance between the centers of the different classes. In this manner, FNPAQR preserves the contribution of samples to different classes [35]. This study is the first to use FNPAQR to classify normal and demented individuals during WM tasks. Finally, the classification staging is necessary to predict the qualitative properties of the mental state of dementia patients. In this stage, the feature vectors extracted from the previous stage are classified into three categories, namely, normal healthy, MCI, and dementia.

In EEG applications, highly accurate classification is related to the quality of extracted features, the reduction in dimensionality, and the classifiers used. Support vector machine (SVM) and k-nearest neighbors (kNN) classifiers are used in biomedical signal classification, such as brain disorder and dementia classification, because of their high accuracy and excel-lent performance [32, 36, 37]. Therefore, this study applies both classifiers.

This study employs ICA–WT denoising technique, the linear and non-linear sets of features, dimensionality reduction technique, SVM and kNN classifiers to obtain EEG valu-able markers and reliable indices for discrimination of VaD and stroke-related MCI patients. Overall, the ultimate goal is to help medical doctors and clinicians in planning and providing a reliable prediction of the course of a disease in addition to the optimal therapeutic program so that dementia patients can experience additional years of high quality of life.

2. Methods and MaterialsIn order to discriminate stroke-related MCI and VaD patients using EEG signal analy-

sis, the recorded EEG signals must go through successive signal processing and analysis stages, these are denoising, feature extraction, dimensionality reduction and dementia classi-fication followed by classifiers performance measurements. Figure 1 presents a block dia-gram of the proposed method.

Figure 1. Block diagram of this study.

2.1 Subjects and Experimental Procedure

EEGs were obtained from 35 participants (5 VaD:2 women and 3 men; age = 64.6±4.8 years, mean ± standard deviation, SD), 15 stroke-related MCI patients (10 women and 5 men; age = 60.26±7.77 years, mean ± standard deviation, SD), and 15 normal records (8 women and 7 men; age = 60.06±5.21 years, mean ± standard deviation, SD) fulfilling the criteria of probable AD. The stroke patients were recruited from the stroke ward of the Pusat Perubatan Universiti Kebangsaan Malaysia (PPUKM), the Medical center of the National University of Malaysia. The VaD patients were recruited from the PPUKM Neurology clinic. The control subjects had no previous history of mental and neurological abnormalities. All control subjects, MCI, and VaD patients underwent cognitive evaluation, including Mini-Mental State Examination [38] and Montreal Cognitive Assessment [39]. The MMSE scores for these VaD and stroke-related MCI patients were 14.8±1.92 and 20.2±5.63 (mean ± SD), respectively. Add to that the MoCA scores for these VaD and stroke-related MCI patients were 13.2±2.38 and 16.13±5.97 (mean ± SD), respectively. The MMSE and MoCA scores for the control subjects were 29.6±0.73 (mean ± SD) and 29.06±0.88 (mean ± SD), respec-tively.

All experiment protocols were approved by the Human Ethics Committee of the Na-tional University of Malaysia. Signed informed consent forms were also obtained from the participants.

The EEG datasets were recorded using the NicoletOne system (V32), which was de-signed and manufactured by VIASYS Healthcare Inc., USA. Nineteen electrodes, as well as ground and system reference electrodes, were positioned on the scalp using cap electrodes according to the 10-20 international system (i.e., Fp2, F8, T4, T6, O2, Fp1, F7, T3, T5, O1, F4, C4, P3, F3, C3, P3, Fz, Cz, and Pz).

The EEG device contains a set of hardware low-pass, high-pass, and notch filters, with frequency values of 0.3 Hz, 70 Hz, and 50 Hz, respectively. Based on the application, the sampling frequency was set to 256 Hz, the impedance of electrode/skin was below 10 kΩ, and the sensitivity was 100 µv/cm. A 12-bit A/D converter was used to digitize the signal and achieve high accuracy.

The session started with a 0.5 second fixation cue when the subjects were asked to be relaxed and avoid the movement activity as much as possible. A simple WM task was then per-formed, during which the subjects were asked to memorize five words for 10 seconds. After-ward, they were asked to remember these words with their eyes closed, and the EEG data were recorded. After 60 seconds, the patients were asked to open their eyes and enumerate all words that they could remember (Figure 2) [40, 41].

Figure 2. The working memory experimental paradigm [41].

2.2 ICA−¿WT Denoising Technique

In this study, the ICA−¿WT denoising technique was used to combine the positive aspects of both ICA and WT and control several of their limitations. However, ICA cannot guarantee that some individual independent components (ICs) contain only noise and do not

contain information about useful EEG channels. Hence, the problem of detection and filtering of ‘‘useful’’ part of each IC is still open, and additional tools are needed to solve it.

Therefore, in order to detect and remove OAs including (eye blinks, vertical and hori-zontal eye movements) and generic discontinuities including (electrode movement and high impedance) from EEG data, the ICA−¿WT denoising technique was used to combine the positive aspects of both ICA and DWT and control several of their limitations. Mathematical details are simplified in the text to focus on the EEG’s role as a marker and a reliable index in detecting the changes that accompany MCI and VaD. The FastICA algorithm proposed by Hyvärinen [42] was used for its simplicity, fast convergence, and efficiency [43]. FastICA decomposes the recorded EEG dataset and extracts new independent component (IC) matri-ces which are marked as artifactual sources. The IC artifact components were marked using ADJUST algorithm [14] that identifies the artifactual ICs on the scalp topographic maps ob-tained by EEGLAB version 13.0.1b [15], as shown in Figure 3.

Figure 3. The scalp map of the projection ICs for the first control subject. IC4, IC6, IC12, IC16 and IC19 account as artifactual components by ADJUST algorithm by EEGLAB

The marked ICs will not be cancelled, but they will be arranged in a new dataset to denoise using DWT. The Symlet mother wavelet of order 9 ‘sym9’ and five decomposition levels were selected to decompose the EEG signals because the sampling frequency of 256 Hz was used in this study [41]. The decomposition coefficients of the six sub-signals through the DWT are cD1, cD2, cD3, cD4, cD5 and cA5, which represent the frequency content of the band-limited EEG signal (where cA is the decomposition approximation coefficient, and cD is the decomposition detail coefficients). The SURE threshold, is an adaptive soft thresh-olding method, which is finding the threshold limit for each level based on Stein’s unbiased risk estimation [44] and commonly used value in [41, 45-47] has been used. Finally, two re-construction steps were performed, namely, the inverse DWT (IDWT) and inverse ICA (inv-ICA). The denoised EEG dataset are used in the next step.

2.3 Feature ExtractionTo develop a non-invasive useful diagnostic index that can be valuable for diagnosing

dementia in the early stages and discriminating the degree of severity, the specifications of system interpretability should be enhanced. Such enhancement can help to improve the clas-sifier performance. This study uses linear and non-linear approaches to meet this demand.

2.3.1 Linear feature extraction

Linear spectral analysis has been intensively used to detect abnormalities in the spec-tra of dementia patients’ EEG. As WT uses a variable window size across the whole signal length, the changes in EEG in different frequency bands can be quantified. The denoised dataset from the previous stage were decomposed using ‘sym9’ into five decomposition lev-els. The decomposition coefficients through the DWT at level 2 to 5 were reconstructed using IDWT. The four reconstruction detail sub-signals (i.e., D2, D3, D4 and D5) and reconstruc-tion approximation sub-signal (A5) provided signal information with respect to each EEG frequency band (Table 1). The five bands of the extracted wavelet coefficients provided a compact representation of the EEG signals, which were submitted for feature extraction anal-ysis stage.

Table 1. The EEG signal decomposition into five frequency bands

Decomposition levels Decomposed Signals EEG bands Frequency bands (Hz)

1 D1 Noises 64-1282 D2 Gamma 32-643 D3 Beta 16-324 D4 Alpha 8-165 D5 Theta 4-85 A5 Delta 0-4

Therefore, linear spectral analysis has been used extensively to detect abnormalities in the spectra of healthy normal, stroke-related MCI, and VaD patients’ EEG dataset. The con-ventional visual characteristics related to dementia can be summarized by slowing the EEG dominant frequency [48, 49]. In the present work, to quantify EEG changes, the relative power (RP) in delta (δRP), theta (θRP), alpha (αRP), beta (βRP), and gamma (γRP) were cal-culated to the WT decomposed signals to distinguish VaD and stroke-related MCI patients EEGs’ from the normal age-match healthy subjects. In order to estimate The RP for each se-lected frequency band δ , θ, α , β, and γ can be calculated using Equation 1 [50]

RP (%)=∑ Selected frequency range

∑Total range (0.5−64 Hz)(1)

In particular, the EEG signals abnormalities in the spectra of VaD and MCI patients can be concluded by slowing the activity of EEG signal in dementia patients that shows a shift in power spectrum toward low frequency compared to healthy age-matched subjects.

Subsequently, the power ratio of ((δ /θ), (θ /α), (α / β), (β /γ) and (θ /γ)) for these spec-tral potentials were calculated. Additionally, this study is intended to be focused on the mark-ers obtained from EEG in order to detect the changes consequent the stroke-related MCI and VaD during WM task.

2.3.2 Non-linear feature extraction

Given that the brain can investigate the complex dynamic information that is reflected from the brain cortex and recorded by EEG devices, this study employed non-linear tech-niques, including PerEn and FD.

A. Permutation entropy (PerEn)Theoretically, PerEn is based on the counting of ordinal patterns (motifs) that describe

the up-and-down changes in the dynamical signal. The concept of PerEn is based on the mea-

sure of the relative frequencies of different motifs. In this study, PerEn is expected to quan-tify the irregularity of EEG signals by discerning the relative change in irregularity from the recorded dataset. PerEn also provides an alternate way of measuring similarity among pat-terns with respect to other types of irregulairty measurements [28].

In order to estimate the PerEn, assume the EEG time series of y= { y1 , y2 ,…, y N } of length N , at each time t of y a vector including the dth subsequence value constructed as: Y t

d ,l= { y t , y t+ 1 ,…, y t+(d−2 ) l , y t+(d−1) l } for t=1,2 ,…, N−(d−1 ) l, where d is the embedded di-mension, determines how much information is contained in each vector and l is the time de-lay. To calculate the PerEn, the d of y i are associated with numbers from 1 to d and arranged in increasing order as { y t+( j1−1 )l , y t+( j2−1 ) l , …, y t+( jd−1−1) l , y t+( jd−1) l } for different samples, there will be d ! Potential ordinal patterns, π, which are named “motifs”. For each π t, p ( π t ) demon-strate the relative frequency as follows:

p (π id , l )=

¿ {t∨t ≤ N−d , type (Y td , l )=π i

d ,l }N−d+1

(2)

where ¿{}denotes the cardinality of the set (the number of elements). The PerEn is computed as follows:

H ( y ,d , l )=− ∑π k=1

πk=d !

p (π k ) ln p ( πk ) (3)

When all motifs have equal probability, the largest value of PerEn is obtained, which has a value of ln d !, where d=3 ,l=1. In contrast, if there is only one p ( πk )different from zero, which illustrates a completely regular signal, the smallest value of PerEn is obtained as much as 0 [51-53]. For 60 seconds, N=15360 samples, 6 windows of 10 second length (2560 sam-ples) were extracted from the original EEG time series for each 19 channels.

B. Fractal dimension (FD)Many algorithms are available to compute FD, like those proposed by Higuchi [54],

Maragos and Sun [55], Katz [56] and Petrosian [57], or the box counting method [58]. Higuchi’s fractal dimension (HFD) is an appropriate method for analysing the FD of biomed-ical signals [59], because of its dependence on a binary sequence and, in many cases, less sensitivity to noise [60]. Given its high computational burden in time and memory [61], box computing is less efficient than HFD. Moreover, Higuchi’s algorithm can estimate FD more accurately than those of Maragos and Sun, Katz, and Petrosian [59, 60]. By contrast, HFD is more sensitive to noise than Katz’s FD [60]. Although the methods of Petrosian and Katz are faster than that of Higuchi, these three algorithms can be run in real time [60]. Therefore, this study applied HFD to examine the EEG background activity in VaD patients, stroke-related MCI patients, and healthy control subjects.

Given a one dimensional time series X=x [ 1 ] , x [2 ] , …, x [ N ], the algorithm to compute the HFD can be described as follows [54]:

X km={x [ m ] , x [ m+k ] , x [m+2k ] , …, x [m+⌊ N−m

k⌋×k ]} (4)

where k and m are integers, k indicates the discrete time interval between points, whereasm=1,2 ,…,k represents the initial time value. For each of the k time series X k

m, the length Lm (k ) can be compute as in equation 5:

Lm (k )={( ∑

i=1

⌊ N −mk ⌋

|x [ m+ik ]−x [m+ (i−1 )× k ]|) ( N−1 )

⌊ N−mk

⌋× k }k

(5)

where N

is the length of the original time series X

and the term

(N−1 )

⌊ N−mk

⌋× k represents the normalization factor. Then, the length of the curve for the time interval k is defined as the av-erage of the k values Lm (k ), for m=1,2 ,…,k:

L (k )=1k

×∑m=1

k

Lm (k ) (6)

Finally, when L (k ) is plotted against 1/kon a double logarithmic scale, with k=1,2 , …,k max, the data should fall on a straight line, with a slope equal to the FD of X . Thus, HFD is defined as the slope of the line that fits the pairs { ln [L (k ) ] , ln (1/k ) } in a least-squares sense. In order to choose an appropriate value of the parameter k max, FD values were plotted against a range of k max. The point at which the FD plateaus is considered a saturation point and that k max value should be selected [62, 63]. In this research, the value of k max=20was chosen and calculated in windows of 10 second length (2560 samples).

3. Statistical AnalysisPreliminarily, the 19 channels from the EEG dataset of the 5 VaD, 15 stroke-related

MCI patients, and 15 normal subjects were grouped into 5 recording regions corresponding to the scalp area of the cerebral cortex. These are the frontal region (seven channels: Fp1, Fp2, F3, F4, F7, F8 and Fz), the temporal region (four channels: T3, T4, T5 and T6), the parietal region (three channels: (P3, P4 and Pz), the occipital region (two channels: O1 and O2), and central region (three channels: C3, C4 and Cz). Next, normality was assessed with Kol-mogorov-Smirnov test, whereas homoscedasticity was verified with Levene’s test. Statistical analysis was performed through ANOVAs in SPSS 22.

In the first session of ANOVA, two-way ANOVA was conducted, the group factor (control healthy subjects, stroke-related MCI patients and VaD patients) was the independent variable and the RP in (δRP, θRP, α RP, βRP, and γRP) was the dependent variable. The sig-nificance was set at p ˂ 0.05. Duncan’s test was used for post-hoc comparison. The signifi-cance for all statistical tests was set at p ˂ 0.05. A second session of ANOVA, two-way ANOVA was performed on the power ratios. The significant differences among the five groups of the scalp regions and ((δ /θ), (θ /α), (α / β), (β /γ) and (θ /γ)) as a dependent variable were evaluated. Post-hoc comparison was performed through Duncan’s test. The significance was set at p˂0.05 .In the third session of ANOVA, two-way ANOVA was applied on the PerEn features. In this analysis, group factor (i.e., control healthy subjects, stroke-related MCI patients, and VaD patients) and the five groups of the scalp regions were the indepen-dent variables, whereas the non-linear feature PerEn was the dependent variable. The fourth session of ANOVA, two-way ANOVA was applied on the FD features. In this analysis, group factor (i.e., control healthy subjects, stroke-related MCI patients, and VaD patients) and the five groups of the scalp regions were the independent variables, whereas the non-lin-ear features FD was the dependent variable. The post-hoc comparison was performed through Duncan’s test. The significance for all statistical tests was set at p ˂ 0.05.

Moreover, Pearson’s correlation coefficient (r) was used to compute correlations be-tween the linear and non-linear features to evaluate the proposed features correlation. The first sessions of Pearson’s correlation were performed on linear features (δ /θ) and (α / β) for the stroke-related MCI patients. In the second session, Pearson’s correlation were calculated on the non-linear feature (PerEn and FD) for VaD and stroke-related MCI patients. The third sessions of Pearson’s correlation were performed on linear and non-linear features for the stroke-related MCI and VaD patients. Statistical significance was set at p ˂0.05 and p ˂0.01for Pearson’s correlations.

4. Preprocessing of Features Before ClassificationIn this application, each EEG channel was characterized with linear features ((δ /θ), (

α / β) and (θ /γ)) and non-linear feature (PerEn and FD). These features were selected on the basis of previous studies that showed their usefulness in distinguishing the EEGs of dementia and MCI patients from those of healthy normal subjects [17, 20, 28, 29, 64].

The features that have been extracted from the previous stage must be analyzed further before they are applied to the classifier. Dimensionality reduction techniques must be employed to avoid problems that are associated with the high dimension feature vector, which may lead to the “curse of dimensionality,” and to reduce the computational time. This study employed dimensionality reduction techniques to avoid classifier overloading, increase the classification model accuracy, and decrease overfitting problems. Therefore, the dimension of feature vectors must be reduced using these techniques.

The dimension of the feature matrix for the healthy control subjects and MCI patients was (90 × 95), where (15 subjects × 6 epochs) = 90 observations and (5 features × 19 channels) = 95 attributes, while that of the feature matrix for the VaD patients was (30 × 95), where (5 VaD × 6 epochs) = 30 observations and (5 features × 19 channels) = 95 attributes. Therefore, VaD is an imbalanced dataset that may affect the performance of the learning algorithms. The main problem of learning from imbalanced datasets is that the imbalance compromises the performance of the learning algorithms. Given that the majority of the learning algorithms assume a balanced class distribution, their results typically favor the predominant class that gives poor class predictions. The class imbalance in the dataset highly affects the quality of the classification model. However, given that the minority class cannot be easily discriminated, the classifier can simply classify each instance as the majority class.

In this study, the minority class was represented by the VaD patients. A synthetic oversampling technique (SMOTE) was applied to overcome the data imbalance [65]. The classifier parameters and percentage of oversampling were determined via 10-fold cross validation using a grid search approach to avoid overfitting and bias in the classification analysis [66]. The available dataset was divided into 10equal size disjoint subsets. One of these subsets was used as the test set, while the remaining nine subsets were combined into a training set to learn the classifier. This procedure was performed 10 times, which resulted in 10 accuracies. The average of these accuracies represented the 10-fold cross-validation accuracy of learning from this dataset [67].

Given that SMOTE changes the dataset, the percentage of oversampling were combined with the parameters. Therefore, those parameters that are found with different SMOTE percentages may not be the same. Using only the training set, the SMOTE was used to equalize the frequency of the classes [68, 69].

This study also employed the fuzzy neighborhood preserving analysis with QR-decomposition (FNPAQR) dimensionality reduction technique [35] of Khushaba et al. [35] to minimize the distance between samples belonging to the same class while maximizing the distance between the centers of different classes. In this manner, FNPAQR preserves the contribution of samples to different classes [35]. This study is the first to use FNPAQR to classify normal and demented individuals during WM tasks.

To project the input features vector using FNPAQR, the projection matrix (GFNDAQR) was initially calculated based on the training set. Afterward, both the training and testing sets were multiplied by this projection matrix to achieve dimensionality reduction. Figure 4 illus-trates the process of projecting the input feature vector for the training data using FNPAQR. Projecting the testing set of the feature vector only involves the multiplication of the testing set with the projection matrix that is calculated on the training.

Moreover, to classify VaD, stroke-related MCI patients and control healthy subjects using SVM and kNN classifiers, a comparative study has been performed to check the effectiveness of using FNPAQR dimensionality reduction technique with the SVM and kNN classifiers in early identification of VaD.

Figure 4. Process of projecting the training and testing features vectors with FNPAQR. GFNDAQR is the projection matrix.

5. Dementia Classification Techniques A final EEG signals analysis was conducted to classify the mentality impairment and

cognition of the subjects into three categories (normal, stroke-related MCI, and VaD). In this final classification stage, the accuracy of classifiers strongly depends on the quality of the ex-tracted features. Therefore, the suitable selection of dimensionality reduction techniques and the type of the classifier can both affect the perfection of the classification results. This study employed two popular methods for brain disorder classification, namely, SVM and kNN [33, 70].5.1 Support Vector Machine (SVM)

Proposed by Vapnik and developed based on computational learning theory [71], SVM has been extensively used in biomedical engineering, particularly in classification, re-gression, and density estimation [72].

Despite being a linear classifier, SVM allows the researcher to generate non-linear classifiers through the non-linear mapping of input patterns into a high dimensional feature space. Therefore, SVM aims to construct hyperplanes that maximize the margin of separation and minimize the misclassification error [73].

5.2 k-nearest Neighbor Classifier (kNN)kNN is among the most popular and simplest non-parametric classification algorithms

in which the classifier labels the samples in the training set based on their similarity The kNN algorithm classifies objects based on the closest training observations that are presented in the significant features matrix [74]. The object is assigned to the class that is most common among its k nearest neighbors, where k is always a positive integer. kNN is more robust when k > 1, and a larger k value can help reduce the effect of noisy points within the training set [75]. The parameter k is generally determined by the characteristics of the datasets, and the nearest samples are assumed to contribute more than the far samples. The unknown sample belongs to the class that is common among the kNN.

6. Performance Measure

The performance of the proposed system was evaluated using average classification accuracy and confusion matrix.

6.1 Average Classification Accuracy

The average classification accuracy of the proposed system was used to characterize the dementia classification results. The average classification accuracy is computed as a per-centage as in Equations 11:

AverageClassification Accuracy= Number of correctly classified instancesTotalnumber of instances

× 100(11)

6.2 Confusion MatrixThe confusion matrix presents another way to plot the performance of the classifica-

tion in terms of accuracy, sensitivity (recall), and precision. Those confusion matrices that show the classification results of the classifiers that are used for classifying the EEG signals are represented by the projected set of features. These matrices represent the frequency with which an EEG segment is misclassified as another. In the confusion matrix for three classes of multi-class classifier task (e.g., A, B, and C), the results on the diagonal are the correct classification accuracies, while those outside the diagonal are the between-class classification errors. For class A, the accuracy, sensitivity (recall), and precision of the classifiers can be calculated as follows using Equations 12 to 14:

Accuracy=TPAA+TN BB+TN CC

Total predicted A+Total predicted B+Total predicted C(12)

Sensitivity=TPAA

TPAA+FN AB+FN AC×100 (13)

Precision=TPAA

TPAA+FP AB+FP AC× 100 (14)

where TP is the true positive, FN is the false negative, FP is the false positive, TN is the true negative. Where Total predicted A=¿ TPAA+FPAB+FPAC , Total predicted B=FN AB+TN BB+FN CB and Total predicted C=FN AC+FNBC+TNCC.

7. Results and Discussion

7.1 Results of EEG DenoisingAfter ICA decomposition was performed using FastICA algorithm and the estimated

ICs were inspected using ADJUST algorithm. The critical selected ICs for the VaD patients, stroke-related MCI patients, and control subjects are summarized in Table 2.

Table 2. Artifactual ICs detected by using ADJUST algorithm by EEGLAB for the VaD, stroke- related MCI patients and control subjects

Dataset Components Dataset Components Dataset ComponentsVaD 1 IC12 MCI 1 IC4, IC6, IC7, IC17 Control 1 IC4, IC6, IC12, IC16,

IC19

2 IC1, IC2, IC3, IC5, IC10 2 IC6 2 IC1, IC2, IC8, IC13,

IC18

3 IC1 3 IC11, IC13, IC14, IC16 3 IC 4, IC12, IC17

4 IC1, IC16 4 -- 4 IC1, IC2

5 IC1 5 IC3, IC6, IC10, IC12 5IC1, IC2, IC3, IC4, IC5, IC6, IC7, IC12, IC15, IC17

6 IC2, IC3 6 IC7, IC10, IC14, IC15, IC17

7 IC1, IC2, IC5, IC17, 7 IC12, IC19

IC18, IC198 IC2, IC5 8 IC1, IC3, IC11, IC149 IC1 9 IC1, IC2

10 IC1, IC3, IC6, IC7, IC18 10 IC1, IC5, IC7, IC13,

IC18, IC1911 IC1, IC13 11 IC1, IC2, IC3, IC14

12

IC1, IC2, IC3, IC4, IC6, IC7, IC8, IC9, IC10, IC12, IC14, IC15, IC16, IC18

12 IC1, IC19

13 -- 13 IC1, IC2, IC3, IC1414 IC1, IC3, IC10 14 IC1, IC2, IC515 IC1, IC6, IC16, IC19 15 IC1, IC2, IC7

Figure 5 shows the ICA−¿WT denoising technique results to channel ‘F8’. It can be observed that the OA artifactual components were sufficiently and successfully suppressed (blue color) compared with the original recorded EEG (dashed black color).

Seconds0 5000 10000 15000

-50

0

50

Ch5

Seconds1200 1400 1600 1800 2000 2200 2400

Mic

rovolts

-40

-20

0

20

Original EEGDenoised EEG

Figure 5. The removal results after the ICA−¿WT technique was applied on the EEG channel F8 to remove OA.

7.2 Results of Statistical Analysis

7.2.1 Linear feature extractionThe statistical characterizing of the differences in linear spectral distributions among

the VaD, stroke-related MCI patients and control healthy subjects will be discussed in the fol-lowing sections. Figure 6 shows the statistical characterizing of the differences in power spectral density dis-tribution among the VaD, stroke-related MCI patients and control healthy subjects. On the one hand, the RP in δ (δRPVaD > δRPMCI > δRPControl) significantly increased and reached the highest values at the frontal and occipital regions ( p < 0.05). θRP significantly increased (θRPVaD > θRPMCI > θRPControl), with their highest values at the occipital, parietal, temporal, and central regions (p < 0.05). γRP significantly increased (γRPVaD > γRPMCI > γRPControl) and reached the highest values at the central, temporal, and frontal regions ( p < 0.05). On the other hand, α power decreased αRPMCI> αRPControl > αRPVaD). αRPwas significantly smaller in the frontal, temporal, and parietal regions ( p < 0.05). Notably, α RPMCI was significantly higher than α RPcontrol. This result could be related to a compensation mechanism in MCI patients dur-

ing memory load and cognitive performance, whereas the healthy control subjects did not have to compensate and the VaD patients could not compensate anymore [76].βRP activity was smaller among the VaD and MCI patients compared with that in the control subjects (βRPVaD< βRPMCI < βRPControl).

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

δ θ α β γ

0

0.1

0.2

0.3

0.4

0.5

0.6

Control Stroke_related MCI VaD

Figure 6. Comparative plot of the relative powers for the five scalp regions of the brain for VaD, Stroke related MCI patients and control subjects.

Figure 7 shows the power ratio statistical characterization of RP spectral density for the subjects belonging to the different categories (control, MCI and VaD patients). Both (δ /θ) and (α / β) of the MCI components are higher compared to the other components. In (δ /θ), interestingly MCI components are significantly higher compared to control subjects this is may be due to memory load as θ is believed to be feasible for cognitive and WM understand-ing [77], so that in the MCI during WM load and cognitive performance is higher whereas the control subjects did not have to compensate and the VaD patients cannot compensate any-more ((δ /θ)MCI > (δ /θ)Control> (δ /θ)VaD). However, the (α / β) of the MCI components is higher than the other components ((α / β)MCI> (α / β)VaD > (α / β)Control), but it is insignificantly differ-entiated among the three groups in all scalp regions. Moreover, (θ /α ) and (θ /γ) are higher in the VaD patients compared to MCI patients and control subjects, but (θ /γ) is significantly differentiate the VaD patients compared to MCI patients and control subjects ¿¿ > (θ /γ )MCI> (θ /γ )Control) in all scalp regions. Notabily, (β /γ) is higher in the control subjects compared to VaD and MCI patients ¿¿ > (β /γ )MCI> (β /γ)VaD) in all scalp regions. It can be concluded that the (δ /θ) and (α / β) significantly shows slowing in the MCI and VaD patients and could be an indicator for MCI patients, whereas the (θ /γ) are markers for VaD detection. Finally, (θ /γ) can be a marker for memory decline in VaD and MCI, and increase with the disease severity [18].To sum up, the (δ /θ) and (α / β) could be the reliable indices that associated with the MCI de-tection whereas (θ /γ) ratios could be considered as reliable indices that associated with the

VaD identification. So far, these EEG markers might be valuable physiological information that help in improve diagnostic procedure.

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

Fron

tal

Tem

pora

l

Pari

etal

Occ

ipit

al

Cent

ral

δ/θ θ/α α/β β/γ θ/γ

0

1

2

3

4

5

6

7

Control Stroke-related MCI VaD

Figure 7. Comparative plot of the power ratios for the five scalp regions of the brain for VaD, Stroke related MCI patients and control subjects.

7.2.2 Non-linear feature extractionThe EEG dataset was successfully denoised using the novel, fully automatic AICA–

WT denoising technique introduced in [12]. The differences in the linear spectral distributions of VaD patients, stroke-related MCI patients, and healthy control subjects in terms of slowing down EEG were statistically characterized in our previous studies [78-80].

Non-linear techniques are widely used in characterizing electrical activities in the brain in terms of EEG complexity. Therefore, in this study, non-linear measures can be applied to analyze EEG datasets of VaD patients, stroke-related MCI patients, and healthy control subjects.

A. Permuatation Entropy (PerEn)PerEn was used to distinguish the brain states of VaD and stroke-related MCI patients

from those of healthy control subjects. Figure 8 presents a comparative plot of PerEn, which is estimated over five scalp regions for the VaD patients, stroke-related MCI patients, and healthy control subjects. The VaD and stroke-related MCI patients had lower PerEn values at five scalp regions (PerEnVaD < PerEnMCI< PerEnControl), and significant differences were observed among the frontal, temporal, and central regions ( p < 0.05). These results suggest that the EEG activity of VaD and stroke-related MCI patients are significantly more regular in the frontal, temporal, and central regions than in a normal brain.

The VaD patients exhibited lower irregularity than the stroke-related MCI patients and healthy control subjects. As expected, the complexity of the EEG signals generally decreases along with the severity of the illness, particularly for the healthy control subjects and the VaD patients.

Frontal Temporal Parietal Occipital Central1.58

1.6

1.62

1.64

1.66

1.68

1.7PerEn Control MCI VaD

Figure 8. Comparative plot of the PerEn for the five scalp regions of the brain for VaD, Stroke related MCI patients and control subjects.

B. Fractal Dimension (FD)FD was proven as an effective candidate in differentiating the VaD and stroke-related

MCI patients from the healthy control subjects. The FD results reveal that EEG signals have less complexity in VaD patients. Figure 9 presents a comparative plot of FD that is estimated over five scalp regions of VaD patients, stroke-related MCI patients, and healthy control sub-jects. The VaD and stroke-related MCI patients had lower FD values at these regions ( FDVaD

< FDMCI< FDControl), and significant differences were observed among the frontal, temporal, and central regions ( p < 0.05). These results suggest that the EEG activity of VaD and stroke-related MCI patients are significantly less complex in these regions than in a normal control brain.

Frontal Temporal Parietal Occipital Central1.55

1.6

1.65

1.7

1.75

1.8

1.85FD Control Stroke-related MCI VaD

Figure 9. Comparative plot of the FD for the five scalp regions of the brain for VaD, Stroke related MCI patients and control subjects.

7.2.3 Correlation analysis

As (δ /θ) and (α / β) could be the reliable indices that associated with the stroke-re-lated MCI detection, the correlation analysis between these linear features for the stroke-re-lated MCI patients are shown in Table 3.

Although the correlations were consistently the strongest between temporal (δ /θ) and the frontal (α / β) (which could reflect the greatest value among the stroke-related MCI pa-tients), most of the other correlations differed significantly among the other scalp regions.

It can be observed that the strongest correlation has appeared between the temporal (δ /θ) and the frontal (α / β). Temporal (δ /θ) and the frontal (α / β) have significant positive correlations and the strength of the association is moderate (r= 0.485, p ˂ 0.01). Thus, con-sidering stroke-related MCI patients, temporal (δ /θ) and the frontal (α / β) are best associated with WM decline (i.e. the higher temporal (δ /θ) and frontal (α / β) correlation characteristic give an indication of reducing the performance of WM tasks. Moreover, this correlations might be interpreted as evidence for a relationship with the frontal and temporal scalp regions particularly memory load in the frontal region and memory transfer to storage in the temporal region.

Table 3. Correlations between (δ /θ) and (α / β) for the MCI patients. Significant group differences are marked with an asterisk(s) ** Correlation is significant at the 0.01 level (2-

tailed), * Correlation is significant at the 0.05 level (2-tailed).MCI

δθ

FrontalMCI

δθ

Temporal

MCIδθ

Parietal

MCIδθ

Occipital

MCIδθ

Central

MCIαβ

Frontalr -0.272** 0.485** -0.152* 0.023 -0.199**p 0.01 0.01 0.041 0.762 0.007

MCIαβ

Temporalr -0.105 -0.354** 0.074 0.059 0.084p 0.159 0.01 0.321 0.435 0.265

MCIαβ

Parietalr -0.088 -0.316** -0.149* 0.206** -0.219**p 0.239 0.05 0.046 0.006 0.003

MCI αβ

Occipitalr 0.331** 0.197** -0.038 -0.104 -0.029p 0.01 0.008 0.617 0.166 0.702

MCIαβ

Centralr -0.085 -0.332** -0.093 0.041 -0.224**p 0.257 0.01 0.213 0.588 0.003

As PerEn and FD suggested that the EEG activities of stroke-related MCI and VaD patients have lower degrees of irregularity and complexity, respectively the correlation analy-sis between these non-linear matrices for VaD and stroke-related MCI patients for the five scalp regions as shown in Table 4.

Although the correlations were consistently the strongest between the temporal PerEn and the temporal FD, (which could reflect the greatest value among the VaD patients), the only correlations differed significantly among the other scalp regions was the central PerEn and the central FD. The specificity of these correlations to temporal and central scalp regions may be linked to perform multiple and complex cognitive functions which are related to memory.

It can be observed that the strongest correlation has appeared between the temporal PerEn and temporal FD. Temporal PerEn and temporal FD have significant positive correla-tions and the strength of the association is moderate (r= 0.778, p ˂ 0.01). Thus, considering VaD patients, temporal PerEn and temporal FD are best associated with WM decline that give an indication of reducing the performance of WM tasks.

Table 4. Correlations between PerEn and FD for the VaD patients. Significant group differ-ences are marked with an asterisk(s) ** Correlation is significant at the 0.01 level (2- tailed),

* Correlation is significant at the 0.05 level (2-tailed).VaD

PerEnFrontal

VaDPerEn

Temporal

VaDPerEn

Parietal

VaDPerEn

Occipital

VaDPerEn

CentralVaDFD

Frontal

r -0.217 -0.104 0.013 -0.502 0.538

p 0.547 0.776 0.972 0.139 0.109

VaDFD

Temporal

r 0.485 0.778** -0.397 0.086 -0.428

p 0.155 0.008 0.256 0.813 0.218

VaDFD

Parietal

r 0.597 0.105 0.215 -0.009 -0.56

p 0.068 0.773 0.551 0.981 0.092

VaDFD

Occipital

r -0.504 -0.535 0.176 0.063 0.28

p 0.137 0.111 0.626 0.864 0.434

`VaDFD

Central

r 0.011 -0.249 0.182 0.59 0.752*

p 0.975 0.487 0.615 0.072 0.012

Furthermore, the PerEn and FD matrices for stroke-related MCI patients for the five scalp regions are shown in Table 5. Although the correlations were consistently the strongest between the central PerEn and the central FD, (which could reflect the greatest value among the stroke-related MCI patients), the correlations between the occipital PerEn and occipital FD and the correlation between the frontal PerEn and frontal FD are considered as interesting correlations. Central PerEn and central FD have significant positive correlations and the strength of the association is very strong (r= 0.915, p ˂ 0.01), occipital PerEn and occipital FD have significant positive correlations and the strength of the association is moderate (r= 0.720, p ˂ 0.01) and frontal PerEn and frontal FD have significant positive correlations and the strength of the association is moderate (r= 0.638, p ˂ 0.01). The specificity of these cor-relations to the central, occipital and frontal scalp regions may be linked to perform multiple and complex cognitive functions which are related to memory.

Table 5. Correlations between PerEn and FD for the stroke-related MCI patients. Significant group differences are marked with an asterisk(s) ** Correlation is significant at the 0.01 level

(2- tailed), * Correlation is significant at the 0.05 level (2-tailed).MCI

PerEnFrontal

MCIPerEn

Temporal

MCIPerEn

Parietal

MCIPerEn

Occipital

MCIPerEn

CentralMCIFD

Frontal

r 0.638** -0.235 0.103 -0.105 0.046

p 0.01 0.211 0.59 0.58 0.811

MCIFD

Temporal

r -0.253 .608** -0.128 0.223 -0.234

p 0.178 0.01 0.501 0.236 0.214

MCIFD

Parietal

r -0.163 0.097 0.458* 0.393* -0.427*

p 0.388 0.612 0.011 0.032 0.018

MCIFD

Occipital

r 0.098 -0.079 -0.045 0.720** 0.064

p 0.607 0.677 0.813 0.01 0.737

MCIFD

Central

r -0.273 -0.277 -0.105 -0.077 0.915**

p 0.144 0.138 0.581 0.686 0.01

Furthermore, the correlation of linear vs non-linear values for the stroke-related MCI and VaD patients are shown in Tables 6 and 7 respectively. In Table 6, for the stroke-related MCI patients, most of the linear ratios are not correlated with the non-linear features among the scalp regions, except of 6 out of 100 correlations have significant negative correlations and the strength of the association are mostly weak and 5 out of 100 correlations have signifi-cant positive correlations and the strength of the association are mostly weak. In Table 7, for the VaD patients, most of the linear θ /γ ratios are not correlated with the PerEn and FD fea-tures, except of 4 out of 100 correlations, 2 of the 4 have significant negative correlations and the strength of the association are mostly moderate and the other 2 correlations have signifi-cant positive correlations and the strength of the association are mostly moderate. Thus, lin-ear and non-linear features complement each other, and will be used together in the next stage to discriminate stroke-related MCI and VaD patients. Table 6. Correlations between non-linear (PerEn and FD) and the linear (delta/theta) features for the stroke-related MCI patients. Significant group differences are marked with an aster-isk(s) ** Correlation is significant at the 0.01 level (2- tailed), * Correlation is significant at

the 0.05 level (2-tailed).

MCIαβ

Frontal

MCIαβ

Temporal

MCIαβ

Parietal

MCIαβ

Occipital

MCIαβ

Central

MCIδθ

Frontal

MCIδθ

Temporal

MCIδθ

Parietal

MCIδθ

Occipital

MCIδθ

CentralMCIPerEnFrontal

r -0.324 0.098 -0.051 -0.233 -0.099 0.339 0.165 0.061 0.329 0.309

p 0.081 0.607 0.791 0.216 0.603 0.067 0.383 0.747 0.076 0.097MCIPerEnTemporal

r 0.196 0.185 0.078 0.354 0.069 0.02 0.058 -0.226 -0.143 -0.167

p 0.298 0.327 0.683 0.055 0.717 0.916 0.759 0.23 0.452 0.377MCIPerEnParietal

r -0.047 -0.069 -0.12 -0.351 -0.152 -0.146 -.380* -0.034 -0.113 0.008

p 0.805 0.715 0.529 0.058 0.422 0.441 0.039 0.86 0.551 0.968MCIPerEnOccipital

r 0.083 0.095 -0.155 0.078 0.088 -0.287 -0.269 0.156 -0.222 -0.105

p 0.662 0.619 0.412 0.681 0.644 0.125 0.151 0.411 0.239 0.581MCIPerEnCentral

r -0.269 0.031 -0.074 -0.236 -0.336 0.171 0.224 0.098 0.266 .399*

p 0.151 0.869 0.699 0.209 0.069 0.365 0.234 0.607 0.155 0.029MCIFDFrontal

r -.469** -0.027 -0.129 -0.218 -0.136 0.213 0.026 -0.007 0.242 .395*

p 0.009 0.886 0.498 0.247 0.473 0.259 0.891 0.972 0.197 0.031MCIFDTemporal

r .393* 0.037 0.07 .561** 0.22 -0.144 -0.1 0.033 -0.315 -0.268

p 0.032 0.845 0.714 0.001 0.242 0.449 0.597 0.861 0.09 0.152MCIFDParietal

r .365* -0.045 0.055 0.108 0.303 -.427* -.442* 0.181 -.578** -.546**

p 0.047 0.813 0.774 0.569 0.103 0.019 0.014 0.338 0.001 0.002MCIFDOccipital

r 0.079 -0.134 -0.283 0.201 -0.121 -0.135 -0.129 0.12 -0.005 0.172

p 0.678 0.482 0.129 0.286 0.524 0.476 0.498 0.526 0.978 0.364MCIFDCentral

r -0.264 -0.017 -0.097 -0.23 -0.312 0.2 0.204 0.138 0.186 0.359

p 0.158 0.93 0.609 0.221 0.094 0.288 0.281 0.466 0.325 0.051

Table 7. Correlations between non-linear (PerEn and FD) and the linear (theta/gamma) fea-tures for the VaD patients. Significant group differences are marked with an asterisk(s) ** Correlation is significant at the 0.01 level (2- tailed), * Correlation is significant at the 0.05

level (2-tailed).VaD

θγ

Frontal

VaDθγ

Temporal

VaDθγ

Parietal

VaDθγ

Occipital

VaDθγ

CentralVaD

PerEnFrontal

r -0.087 -0.067 -0.238 -.686* -0.003

p 0.812 0.854 0.508 0.028 0.993

VaDPerEn

Temporal

r 0.061 -0.348 0.401 -0.552 0.551

p 0.867 0.324 0.25 0.098 0.099

VaDPerEn

Parietal

r -0.2 0.356 -0.219 0.116 -0.256

p 0.58 0.313 0.543 0.75 0.475

VaDPerEn

Occipital

r -0.113 -0.443 -0.199 -0.131 -0.046

p 0.755 0.2 0.582 0.718 0.9

VaDPerEn

Central

r -0.26 -0.512 -0.212 0.589 -0.224

p 0.468 0.13 0.557 0.073 0.534

VaDFD

Frontal

r -0.425 0.048 0.124 0.632* 0.024

p 0.221 0.895 0.733 0.05 0.947

VaDFD

Temporal

r 0.478 -0.014 0.078 -0.799** 0.13

p 0.163 0.969 0.831 0.006 0.721

VaDFD

Parietal

r 0.139 0.136 -0.215 -0.543 -0.07

p 0.701 0.707 0.552 0.105 0.847

VaDFD

Occipital

r -0.275 0.209 -0.11 0.760* -0.214

p 0.441 0.561 0.762 0.011 0.552

VaDFD

Central

r 0.205 -0.494 -0.565 0.188 -0.572

p 0.571 0.147 0.089 0.603 0.084

7.3 Results of Dementia Classification TechniquesThis study also dealt with three-class EEG signals or the multi-class classification

problem. Two classifiers, namely, SVM and kNN, were proposed for classifying the EEG sig-nals of VaD patients, stroke-related MCI patients, and healthy control subjects. These classi-fiers were selected because of their dependence on the sizes of the training and test sets. These classifiers were trained by the same training dataset and tested on the testing dataset to evalu-ate their performance.

We obtained the best results for the SVM classifier by performing a 10-fold cross-validation for optimizing C on the training set. Specifically, the SVMs were trained for different complexity parameter C values with a range of −4 ≤ log10 (C )≤ 4in C values C∈ {0.0001,0.001,0 .01,0 .1,0,10,100,1000 ,10000 }.The best result was obtained for C = 10 in the testing procedure. The RBF-kernel functions were used for implementing the multi-class SVMs classifier. In the SVM training, the smoothing parameter σ was determined based on the minimum misclassification rate that was computed from the training dataset. The optimal σ can only be found by systematically varying its value in different training sessions.

Therefore, the σ value was varied between 0.1 and 1 at intervals of 0.1. The minimum-misclassification rate was attained at σ = 0.5.

The same processing chain has been applied to the linear features and to the non-linear features. SVM achieved classification accuracies of 70.95% using linear spectral features and 87.14% using non-linear features.

Additionally, the same processing analysis has been performed without using FNPAQR dimensionality reduction technique, the SVM achieved classification accuracies of 70%.

The SVM overall performance, using the features set after using FNPAQR, is improved to a 91.48% accuracy (sensitivity of 91.48% and precision of 91.40%) for SVM. This shows the potential of using linear and non-linear methods with FNPAQR to obtain more accurate early detection of VaD rather than using individual method.

The confusion matrix and the average precision and sensitivity and the overall accuracy of the classifications using the combination approach are presented in Table 8.

Table 8. The table of confusion matrix calculations for multi-class classification using FN-PAQR and SVM

Confusion Matrix

VaD MCI Control

VaD 94.44%

5.56% 0%

MCI 8.89% 86.67%

4.44%

Control 0% 6.68% 93.33%

From Table 8, the three diagonal cells show the percentage of correct classification using SVM classifier. For example VaD are correctly classified with 94.44%. Similarly, 86.67% are correctly classified as stroke-related MCI patients and 93.33% are correctly classified as control subjects. 5.56% of VaD patients are incorrectly classified as stroke-related MCI patients. Similarly, 13.33% of stroke-related MCI patients are incorrectly classified as 4.44% control subjects and 8.89% VaD patients. Moreover, 6.68% of control subjects are incorrectly classified as stroke-related MCI patients. For the VaD patients, the precision is 91.40% and sensitivity is 94.44%. Which means that for precision, out of the times that VaD patients were predicted, 91.40% of the time the system was in fact correct. And for sensitivity, it means that out of all the times that VaD patients should have been correctly detected only 94.44% of the other classes were correctly detected.

For the stroke-related MCI patients, the precision is 87.64% and sensitivity is 86.67%, which means that for precision, out of the times that stroke-related MCI patients were predicted, 87.64% of the time the system was in fact correct. And for sensitivity, it means that out of all the times that stroke-related MCI patients should have been correctly detected only 86.67% of the other classes were correctly detected.

For the control subjects, the precision is 95.45% and sensitivity is 93.33%, which means that for precision, out of the times that control subjects were predicted, 95.45% of the time the system was in fact correct. And for sensitivity, it means that out of all the times that control subjects should have been correctly detected only 93.33% of the other classes were correctly detected.

Finally, the overall sensitivity is 91.48%% whereas the overall precision is 91.40%. The results show that these VaD can be differentiated with a high accuracy of 91.48% using SVM classifier.

For kNN classifier, in this study, to select the k value for the kNN classifier, k was varied between 1 and 10 at intervals of 1. The classifier was trained to find the best value of k ,

which was obtained at k = 7, and to maximize the classification accuracy. We used the Euclidean distance as a similarity measure to classify each trial by kNN.

kNN achieved classification accuracies of 66.67% using linear spectral features and 84.29% using non-linear features.

The same processing analysis has been applied without using FNPAQR dimensionality reduction technique, the kNN achieved classification accuracy of 67.78%.

The kNN overall performance, using the features set after using FNPAQR, is improved to 89.63% accuracy (sensitivity of 89.63% and precision of 89.77%) for kNN. This shows the potential of using linear and non-linear methods with FNPAQR to obtain more accurate early detection of VaD rather than using individual method.

The confusion matrix and the average precision and sensitivity and the overall accuracy of the classifications using the combination approach is presented in Table 9.

Table 9. The table of confusion matrix calculations for multi-class classification using FNPAQR and kNN

Confusion Matrix

VaD MCI Control

VaD 92.22%

7.78% 0%

MCI 8.89% 86.67%

4.44%

Control 3.33% 6.67% 90%

From Table 9, the three diagonal cells show the percentage of correct classification using kNN classifier. For example VaD are correctly classified with 92.22%. Similarly, 86.67% are correctly classified as stroke-related MCI patients and 90% are correctly classified as control subjects. 7.78% of VaD patients are incorrectly classified as stroke-related MCI patients. Similarly, 13.33% of stroke-related MCI patients are incorrectly classified as 4.44% control subjects and 8.89% VaD patients. Moreover, 10% of control subjects are incorrectly classified as 3.33% VaD patients and 6.67% stroke-related MCI patients.

For the VaD patients, the precision is 88.30% and sensitivity is 92.22%. Which means that for precision, out of the times that VaD patients were predicted, 88.30% of the time the system was in fact correct. And for sensitivity, it means that out of all the times that VaD patients should have been correctly detected only 92.22% of the other classes were correctly detected.

For the stroke-related MCI patients, the precision is 85.71% and sensitivity is 86.67%, which means that for precision, out of the times that stroke-related MCI patients were predicted, 85.71% of the time the system was in fact correct. And for sensitivity, it means that out of all the times that stroke-related MCI patients should have been correctly detected only 86.67% of the other classes were correctly detected.

For the control subjects, the precision is 95.29% and sensitivity is 90%, which means that for precision, out of the times that control subjects were predicted, 95.29% of the time the system was in fact correct. And for sensitivity, it means that out of all the times that control subjects should have been correctly detected only 90% of the other classes were correctly detected.

Finally, the overall sensitivity is 89.63% whereas the overall precision is 89.77%. The results show that these VaD can be differentiated with a high accuracy of 89.63% using kNN classifier.

In this study, SVM and kNN have been used to classify VaD patients, stroke-related MCI patients, and healthy control subjects. However, kNN obtained slightly lower accuracy than the multi-class SVM. Thus, SVM was included in the study as a benchmark technique as

well as being simple and proven to show good results. SVM supports multi-class classification, to discriminate VaD, stroke-related MCI patients and healthy control subjects’. These results indicated the crucial role of the proposed ICA–WT denoising technique, the representation of signals by the used features, and the FNPAQR dimensionality reduction technique all significantly increase the classification accuracies of these classifiers.

8. Conclusion

The ICA–WT technique is considered as a hybrid technique to denoise the EEG datasets of 5 VaD patients, 15 stroke-related MCI patients, and 15 control subjects during WM tasks. A total of 19 channels from various regions of the scalp were recorded from control subjects, stroke-related MCI patients, and VaD patients during WM tasks. Linear spectral and non-linear analyses were used to quantify changes in the EEG background activities of VaD and stroke-related MCI patients. Linear spectral features including RP and power ratio were calculated to quantify the slowness of the EEG in VaD and stroke-related MCI patients compared to healthy control subjects ( p < 0.05). Non-linear features including PerEn and FD suggested that the EEG activities of stroke-related MCI and VaD patients have lower degrees of irregularity and complexity, respectively. These patients have significantly lower PerEn and FD values than the healthy control subjects ( p < 0.05). Therefore, those brains that are affected by post-stroke dementia show a more regular and less complex EEG behavior in the tested regions. Temporal (δ /θ) and the frontal (α / β) have significant positive correlations and the strength of the association is moderate for stroke-related MCI patients, this gives an indication that (δ /θ) and (α / β) could represent the most sensitive EEG reliable indices of MCI detection, whereas temporal PerEn and temporal FD have significant positive correlations and the strength of the association is moderate for VaD patients. Central PerEn and central FD have significant positive correlations and the strength of the association is very strong for stroke-related MCI patients. Most of the linear ratios are not correlated with the non-linear features among the scalp regions for stroke-related MCI and VaD patients. Thus, linear and non-linear features complement each other, and have used together to discriminate stroke-related MCI and VaD patients. Additionally, the extracted features were applied to the FNPAQR dimensionality reduction technique. SVM and kNN with a 10-fold cross-validation procedure were used. SVM and kNN achieved classification accuracies of 70.95% and 66.67% for linear spectral features and 87.14% and 84.29% for non-linear features, respectively. kNN achieved classification accuracies of 70% and 67.78% without using FNPAQR dimensionality reduction technique respectively, 91.8% and 89.63 % using FNPAQR dimensionality reduction technique, respectively. Then, the linear spectral and non-linear features have been combined, SVM and kNN obtained total classification accuracies of 91.48% and 89.63% respectively, which gave a clear indication that the combined use of the ICA–WT denoising technique, linear and non-linear features, FNPAQR dimensionality reduction technique, and SVM can accurately detect and provide a highly reliable index for identifying and classifying patients with VaD and stroke-related MCI. This study has certain limitations which have to be pointed out. First of all, the sample size was small. Moreover, the patients follow-up was limited. Besides, larger series with long-term follow-up are needed. Despite these drawbacks, all our results are consistent with those of other researchers, whose findings showed that the earliest changes in EEG signals among the dementia patients are related to the increase in δRP, θRP, and γRP activities, as well as the decrease in αRP and βRP activities. Moreover, all our finding confirms the theories that a diffuse irregularity of background activity may be observed in the EEGs of patients with dementia and that the dynamic processes underlying the EEG recording are less irregular for

dementia patients than for normal subjects [16, 22, 28, 49, 77, 79, 81, 82]. Therefore, the ICA–WT denoising technique, the linear and non-linear sets of features, FNPAQR dimensionality reduction technique, SVM classifier can yield useful information for characterizing and identifying valuable EEG markers and reliable indices that are associated with the stroke-related MCI and VaD patients.

AcknowledgmentsThe authors wish to express their gratitude to Mrs. Musmarlina Omar who recruited the

healthy control subjects, Mr. Mohd Izhar Ariff and Neurology Unit staff from the Neurology Unit at PPUKM for their assistance in the acquisition of the EEG brain signals during work-ing memory task. My sincere thanks also goes to Ms. Khairiyah Mohamad from the Neurol-ogy Unit at PPUKM who provided the neuropsychological assessment for all subjects.

Conflicts of Interest

The authors declare no conflict of interest.

References

[1] N. K. Al-Qazzaz, S. H. Ali, S. A. Ahmad, and S. Islam, "Cognitive assessments for the early diagnosis of dementia after stroke," Neuropsychiatric disease and treatment, vol. 10, p. 1743, 2014.

[2] B. Cullen, B. O’Neill, J. J. Evans, R. F. Coen, and B. A. Lawlor, "A review of screening tests for cognitive impairment," Journal of Neurology, Neurosurgery & Psychiatry, vol. 78, pp. 790-799, 2007.

[3] C. McVeigh and P. Passmore, "Vascular dementia: prevention and treatment," Clinical interventions in aging, vol. 1, p. 229, 2006.

[4] A. D. Korczyn, V. Vakhapova, and L. T. Grinberg, "Vascular dementia," J Neurol Sci, vol. 322, pp. 2-10, Nov 15 2012.

[5] N. K. Al-Qazzaz, S. H. B. Ali, S. A. Ahmad, K. Chellappan, M. S. Islam, and J. Escudero, "Role of EEG as Biomarker in the Early Detection and Classification of Dementia," The Scientific World Journal, vol. 2014, 2014.

[6] A. Baddeley, "Working memory," Science, vol. 255, pp. 556-559, 1992.[7] K. Chellappan, N. K. Mohsin, S. Bin Md Ali, and M. Islam, "Post-stroke brain memory assessment

framework," in Biomedical Engineering and Sciences (IECBES), 2012 IEEE EMBS Conference on, 2012, pp. 189-194.

[8] C. Guerrero-Mosquera, A. M. Trigueros, and a. A. Navia-Vazquez, EEG Signal Processing for Epilepsy 2012.

[9] I. M. B. Núñez, "EEG artifact dtection," 2010.[10] R. Sameni and C. Gouy-Pailler, "An iterative subspace denoising algorithm for removing

electroencephalogram ocular artifacts," Journal of neuroscience methods, vol. 225, pp. 97-105, 2014.[11] E. D. Übeyli, "Combined neural network model employing wavelet coefficients for EEG signals

classification," Digital Signal Processing, vol. 19, pp. 297-308, 2009.[12] N. K. Al-Qazzaz, S. Hamid Bin Mohd Ali, S. A. Ahmad, M. S. Islam, and J. Escudero, "Automatic

Artifact Removal in EEG of Normal and Demented Individuals Using ICA–WT during Working Memory Tasks," Sensors, vol. 17, p. 1326, 2017.

[13] N. Mammone, F. La Foresta, and F. C. Morabito, "Automatic artifact rejection from multichannel scalp EEG by wavelet ICA," Sensors Journal, IEEE, vol. 12, pp. 533-542, 2012.

[14] A. Mognon, J. Jovicich, L. Bruzzone, and M. Buiatti, "ADJUST: An automatic EEG artifact detector based on the joint use of spatial and temporal features," Psychophysiology, vol. 48, pp. 229-240, 2011.

[15] A. Delorme and S. Makeig, "EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis," Journal of neuroscience methods, vol. 134, pp. 9-21, 2004.

[16] L. A. Coben, W. Danziger, and M. Storandt, "A longitudinal EEG study of mild senile dementia of Alzheimer type: changes at 1 year and at 2.5 years," Electroencephalography and clinical neurophysiology, vol. 61, pp. 101-112, 1985.

[17] D. Moretti, O. Zanetti, G. Binetti, and G. Frisoni, "Quantitative EEG markers in mild cognitive impairment: degenerative versus vascular brain impairment," International journal of Alzheimer's disease, vol. 2012, 2012.

[18] D. Moretti, C. Fracassi, M. Pievani, C. Geroldi, G. Binetti, O. Zanetti , et al., "Increase of theta/gamma ratio is associated with memory impairment," Clinical Neurophysiology, vol. 120, pp. 295-303, 2009.

[19] D. Moretti, D. Paternicò, G. Binetti, O. Zanetti, and G. Frisoni, "Relationship between EEG alpha3/alpha2 Ration and the nucleus accumbens in subjects with mild cognitive impairment," J Neurol Neurophysiol, vol. 4, pp. 1-6, 2013.

[20] D. Moretti, C. Miniussi, G. Frisoni, O. Zanetti, G. Binetti, C. Geroldi, et al., "Vascular damage and EEG markers in subjects with mild cognitive impairment," Clinical neurophysiology, vol. 118, pp. 1866-1876, 2007.

[21] M. T. Schmidt, P. A. Kanda, L. F. Basile, H. F. da Silva Lopes, R. Baratho, J. Demario , et al., "Index of alpha/theta ratio of the electroencephalogram: a new marker for Alzheimer’s disease," Front Aging Neurosci, vol. 5, p. 60, 2013.

[22] M. Lundqvist, P. Herman, and A. Lansner, "Theta and gamma power increases and alpha/beta power decreases with memory load in an attractor network model," Journal of cognitive neuroscience, vol. 23, pp. 3008-3020, 2011.

[23] D. Abásolo, R. Hornero, C. Gómez, M. García, and M. López, "Analysis of EEG background activity in Alzheimer's disease patients with Lempel–Ziv complexity and central tendency measure," Medical engineering & physics, vol. 28, pp. 315-322, 2006.

[24] J. Escudero, D. Abásolo, R. Hornero, P. Espino, and M. López, "Analysis of electroencephalograms in Alzheimer's disease patients with multiscale entropy," Physiological measurement, vol. 27, p. 1091, 2006.

[25] H. Kantz, J. Kurths, and G. Mayer-Kress, Nonlinear analysis of physiological data: Springer Science & Business Media, 2012.

[26] S. Sanei, Adaptive processing of brain signals: John Wiley & Sons, 2013.[27] A. Holzinger, M. Hörtenhuber, C. Mayer, M. Bachler, S. Wassertheurer, A. J. Pinho , et al., "On

entropy-based data mining," in Interactive Knowledge Discovery and Data Mining in Biomedical Informatics, ed: Springer, 2014, pp. 209-226.

[28] F. C. Morabito, D. Labate, F. La Foresta, A. Bramanti, G. Morabito, and I. Palamara, "Multivariate multi-scale permutation entropy for complexity analysis of Alzheimer’s disease EEG," Entropy, vol. 14, pp. 1186-1202, 2012.

[29] G. Henderson, E. Ifeachor, N. Hudson, C. Goh, N. Outram, S. Wimalaratna, et al., "Development and assessment of methods for detecting dementia using the human electroencephalogram," Biomedical Engineering, IEEE Transactions on, vol. 53, pp. 1557-1568, 2006.

[30] A. C. Yang, S.-J. Wang, K.-L. Lai, C.-F. Tsai, C.-H. Yang, J.-P. Hwang, et al., "Cognitive and neuropsychiatric correlates of EEG dynamic complexity in patients with Alzheimer's disease," Progress in Neuro-Psychopharmacology and Biological Psychiatry, vol. 47, pp. 52-61, 2013.

[31] A. Subasi and M. Ismail Gursoy, "EEG signal classification using PCA, ICA, LDA and support vector machines," Expert Systems with Applications, vol. 37, pp. 8659-8666, 2010.

[32] M. KavitaMahajan and M. S. M. Rajput, "A Comparative study of ANN and SVM for EEG Classification," International Journal of Engineering, vol. 1, 2012.

[33] F. Vialatte, A. Cichocki, G. Dreyfus, T. Musha, T. M. Rutkowski, and R. Gervais, "Blind source separation and sparse bump modelling of time frequency representation of eeg signals: New tools for early detection of alzheimer’s disease," in Machine Learning for Signal Processing, 2005 IEEE Workshop on, 2005, pp. 27-32.

[34] M. Besserve, K. Jerbi, F. Laurent, S. Baillet, J. Martinerie, and L. Garnero, "Classification methods for ongoing EEG and MEG signals," Biological research, vol. 40, pp. 415-437, 2007.

[35] R. N. Khushaba, S. Kodagoda, D. Liu, and G. Dissanayake, "Electromyogram (EMG) based fingers movement recognition using neighborhood preserving analysis with QR-decomposition," in Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), 2011 Seventh International Conference on, 2011, pp. 1-105.

[36] R. H. Chowdhury, M. B. Reaz, M. A. B. M. Ali, A. A. Bakar, K. Chellappan, and T. Chang, "Surface Electromyography Signal Processing and Classification Techniques," Sensors, vol. 13, pp. 12431-12466, 2013.

[37] A. Al-Timemy, G. Bugmann, J. Escudero, and N. Outram, "Classification of Finger Movements for the Dexterous Hand Prosthesis Control with Surface Electromyography," 2013.

[38] M. F. Folstein, S. E. Folstein, and P. R. McHugh, "Mini-mental state," A prac-32, 1998.[39] T. Smith, N. Gildeh, and C. Holmes, "The Montreal Cognitive Assessment: validity and utility in a

memory clinic setting," Canadian Journal of Psychiatry, vol. 52, p. 329, 2007.[40] N. K. Al-Qazzaz, S. H. Ali, S. A. Ahmad, S. Islam, and K. Mohamad, "Cognitive impairment and

memory dysfunction after a stroke diagnosis: a post-stroke memory assessment," Neuropsychiatric disease and treatment, vol. 10, p. 1677, 2014.

[41] N. Al-Qazzaz, S. Hamid Bin Mohd Ali, S. Ahmad, M. Islam, and J. Escudero, "Selection of Mother Wavelet Functions for Multi-Channel EEG Signal Analysis during a Working Memory Task," Sensors, vol. 15, p. 29015, 2015.

[42] A. Hyvarinen, "Fast and robust fixed-point algorithms for independent component analysis," Neural Networks, IEEE Transactions on, vol. 10, pp. 626-634, 1999.

[43] A. H. Al-Timemy, G. Bugmann, N. Outram, and J. Escudero, "Reduction in classification errors for myoelectric control of hand movements with independent component analysis," in The 5th International Conference on Information Technology, ICIT, 2011.

[44] C. M. Stein, "Estimation of the mean of a multivariate normal distribution," The annals of Statistics, pp. 1135-1151, 1981.

[45] R. Romo-Vazquez, R. Ranta, V. Louis-Dorr, and D. Maquin, "EEG ocular artefacts and noise removal," in Engineering in Medicine and Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the IEEE, 2007, pp. 5445-5448.

[46] E. Estrada, H. Nazeran, G. Sierra, F. Ebrahimi, and S. K. Setarehdan, "Wavelet-based EEG denoising for automatic sleep stage classification," in Electrical Communications and Computers (CONIELECOMP), 2011 21st International Conference on, 2011, pp. 295-298.

[47] N. K. Al-Qazzaz, S. Ali, S. A. Ahmad, M. S. Islam, and M. I. Ariff, "Selection of mother wavelets thresholding methods in denoising multi-channel EEG signals during working memory task," in Biomedical Engineering and Sciences (IECBES), 2014 IEEE Conference on, 2014, pp. 214-219.

[48] I. Rosen, "Electroencephalography as a diagnostic tool in dementia," Dementia and geriatric cognitive disorders, vol. 8, pp. 110-116, 1997.

[49] J. Jeong, "EEG dynamics in patients with Alzheimer's disease," Clinical neurophysiology, vol. 115, pp. 1490-1505, 2004.

[50] K.-E. Ko, H.-C. Yang, and K.-B. Sim, "Emotion recognition using EEG signals with relative power values and Bayesian network," International Journal of Control, Automation and Systems, vol. 7, pp. 865-870, 2009.

[51] C. Bandt and B. Pompe, "Permutation entropy: a natural complexity measure for time series," Physical review letters, vol. 88, p. 174102, 2002.

[52] M. Zanin, L. Zunino, O. A. Rosso, and D. Papo, "Permutation entropy and its main biomedical and econophysics applications: a review," Entropy, vol. 14, pp. 1553-1577, 2012.

[53] H. Azami and J. Escudero, "Improved multiscale permutation entropy for biomedical signal analysis: Interpretation and application to electroencephalogram recordings," Biomedical Signal Processing and Control, vol. 23, pp. 28-41, 2016.

[54] T. Higuchi, "Approach to an irregular time series on the basis of the fractal theory," Physica D: Nonlinear Phenomena, vol. 31, pp. 277-283, 1988.

[55] P. Maragos and F.-K. Sun, "Measuring the fractal dimension of signals: morphological covers and iterative optimization," IEEE Transactions on signal Processing, vol. 41, pp. 108-121, 1993.

[56] M. J. Katz, "Fractals and the analysis of waveforms," Computers in biology and medicine, vol. 18, pp. 145-156, 1988.

[57] A. Petrosian, "Kolmogorov complexity of finite sequences and recognition of different preictal EEG patterns," in Computer-Based Medical Systems, 1995., Proceedings of the Eighth IEEE Symposium on, 1995, pp. 212-217.

[58] B. Mandelbrot, "Fractals Form, Chance and Dimension. San Francisco, 1977," ed: WH Freeman and Co.

[59] A. Accardo, M. Affinito, M. Carrozzi, and F. Bouquet, "Use of the fractal dimension for the analysis of electroencephalographic time series," Biological cybernetics, vol. 77, pp. 339-350, 1997.

[60] R. Esteller, G. Vachtsevanos, J. Echauz, and B. Litt, "A comparison of waveform fractal dimension algorithms," Circuits and Systems I: Fundamental Theory and Applications, IEEE Transactions on, vol. 48, pp. 177-183, 2001.

[61] A. Block, W. Von Bloh, and H. Schellnhuber, "Efficient box-counting determination of generalized fractal dimensions," Physical Review A, vol. 42, p. 1869, 1990.

[62] T. L. Doyle, E. L. Dugan, B. Humphries, and R. U. Newton, "Discriminating between elderly and young using a fractal dimension analysis of centre of pressure," International journal of medical sciences, vol. 1, p. 11, 2004.

[63] W. Klonowski, E. Olejarczyk, and R. Stepien, "‘Epileptic seizures’ in economic organism," Physica A: Statistical Mechanics and its Applications, vol. 342, pp. 701-707, 2004.

[64] D. Moretti, G. Frisoni, C. Fracassi, M. Pievani, C. Geroldi, G. Binetti, et al., "MCI patients’ EEGs show group differences between those who progress and those who do not progress to AD," Neurobiology of aging, vol. 32, pp. 563-571, 2011.

[65] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, "SMOTE: synthetic minority over-sampling technique," Journal of artificial intelligence research, pp. 321-357, 2002.

[66] R. Kohavi, "A study of cross-validation and bootstrap for accuracy estimation and model selection," in Ijcai, 1995, pp. 1137-1145.

[67] Y. Song and J. Zhang, "Discriminating preictal and interictal brain states in intracranial EEG by sample entropy and extreme learning machine," Journal of neuroscience methods, vol. 257, pp. 45-54, 2016.

[68] I. H. Witten and E. Frank, Data Mining: Practical machine learning tools and techniques: Morgan Kaufmann, 2005.

[69] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H. Witten, "The WEKA data mining software: an update," ACM SIGKDD explorations newsletter, vol. 11, pp. 10-18, 2009.

[70] C. Lehmann, T. Koenig, V. Jelic, L. Prichep, R. E. John, L.-O. Wahlund, et al., "Application and comparison of classification algorithms for recognition of Alzheimer's disease in electrical brain activity (EEG)," Journal of neuroscience methods, vol. 161, pp. 342-350, 2007.

[71] V. Vapnik, The nature of statistical learning theory: springer, 2000.[72] I. Guler and E. D. Ubeyli, "Multiclass support vector machines for EEG-signals classification,"

Information Technology in Biomedicine, IEEE Transactions on, vol. 11, pp. 117-126, 2007.[73] D. Garrett, D. A. Peterson, C. W. Anderson, and M. H. Thaut, "Comparison of linear, nonlinear, and

feature selection methods for EEG signal classification," Neural Systems and Rehabilitation Engineering, IEEE Transactions on, vol. 11, pp. 141-144, 2003.

[74] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern classification: John Wiley & Sons, 2012.[75] M. Kantardzic, Data mining: concepts, models, methods, and algorithms: John Wiley & Sons, 2011.[76] Y. Pijnenburg, Y. Vd Made, A. V. C. Van Walsum, D. Knol, P. Scheltens, and C. Stam, "EEG

synchronization likelihood in mild cognitive impairment and Alzheimer's disease during a working memory task," Clinical Neurophysiology, vol. 115, pp. 1332-1339, 2004.

[77] W. Klimesch, "EEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis," Brain research reviews, vol. 29, pp. 169-195, 1999.

[78] N. K. Al-Qazzaz, S. Ali, M. S. Islam, S. A. Ahmad, and J. Escudero, "EEG markers for early detection and characterization of vascular dementia during working memory tasks," in Biomedical Engineering and Sciences (IECBES), 2016 IEEE EMBS Conference on, 2016, pp. 347-351.

[79] N. K. Al-Qazzaz, S. Ali, S. Islam, S. Ahmad, and J. Escudero, "EEG Wavelet Spectral Analysis During a Working Memory Tasks in Stroke-Related Mild Cognitive Impairment Patients," in International Conference for Innovation in Biomedical Engineering and Life Sciences, 2016, pp. 82-85.

[80] N. K. Al-Qazzaz, S. Ali, S. A. Ahmad, M. S. Islam, and J. Escudero, "Entropy-Based Markers of EEG Background Activity of Stroke-Related Mild Cognitive Impairment and Vascular Dementia Patients," presented at the 2nd International Conference on Sensors Engineering and Electronics Instrumental Advances (SEIA' 2016),Barcelona, Spain, 2016.

[81] A. Gevins, M. E. Smith, L. McEvoy, and D. Yu, "High-resolution EEG mapping of cortical activation related to working memory: effects of task difficulty, type of processing, and practice," Cerebral cortex, vol. 7, pp. 374-385, 1997.

[82] C. Gómez, Á. Mediavilla, R. Hornero, D. Abásolo, and A. Fernández, "Use of the Higuchi's fractal dimension for the analysis of MEG recordings from Alzheimer's disease patients," Medical engineering & physics, vol. 31, pp. 306-313, 2009.