Working Memory Classification Enhancement of EEG Activity in Dementia: A Comparative Study

The purpose of the current investigation is to distinguish between working memory ( 𝑊𝑀 ) in five patients with vascular dementia ( 𝑉𝐷 ), fifteen post-stroke patients with mild cognitive impairment ( 𝑆𝑀𝐶𝐼 ), and fifteen healthy control individuals ( 𝐻𝐶 ) based on background electroencephalography (EEG) activity. The elimination of EEG artifacts using wavelet (WT) pre-processing denoising is demonstrated in this study. In the current study, spectral entropy ( 𝑆𝑝𝑒𝑐𝐸𝑛 ), permutation entropy ( 𝑃𝑒𝑟𝐸𝑛 ), and approximation entropy ( 𝐴𝑝𝐸𝑛 ) were all explored. To improve the 𝑊𝑀 classification using the k-nearest neighbors ( 𝑘 NN) classifier scheme, a comparative study of using fuzzy neighbourhood preserving analysis with 𝑄𝑅 -decomposition ( 𝐹𝑁𝑃𝐴𝑄𝑅 ) as a dimensionality reduction technique and the improved binary gravitation search ( 𝐼𝐵𝐺𝑆𝐴 ) optimization algorithm as a channel selection method has been conducted. The 𝑘 NN classification accuracy was increased from 86.67% to 88.09% and 90.52% using the 𝐹𝑁𝑃𝐴𝑄𝑅 dimensionality reduction technique and the 𝐼𝐵𝐺𝑆𝐴 channel selection algorithm, respectively. According to the findings, 𝐼𝐵𝐺𝑆𝐴 reliably enhances 𝑊𝑀 discrimination of 𝐻𝐶 , 𝑆𝑀𝐶𝐼 , and 𝑉𝐷 participants. Therefore, WT, entropy features, IBGSA and 𝑘 NN classifiers provide a valid dementia index for looking at EEG background activity in patients with 𝑉𝐷 and 𝑆𝑀𝐶𝐼 .


Introduction
Stroke is broadly categorized by its type and severity and the brain region affected.Dementia can be brought on after a stroke, and its severity relies on how quickly the condition is identified and treated [1].Following a stroke, working memory () impairment is common.Within the first year of stroke diagnosis, vascular dementia () may develop in 30% of stroke patients.The prevalence of  doubles every 5-10 years after the age of 65 in the elderly population.Clinically speaking, mild cognitive impairment (  ), in particular in attention, memory, language and orientation is a transition stage of cognitive decline [2].At the time of diagnosis, attention, executive functioning, and memory show the greatest impact of a stroke [3].
Patients who had suffered from cognitive impairment as a result of a stroke were the first to be introduced to the vascular cognitive impairment (VCI) spectrum, which spans the range from mild cognitive impairment () to advanced dementia.However, the phrase "cognitive impairment no dementia" (CIND) is used to refer to the period of time following dementia during which the brain is in danger [4].People who have  , have a more severe decline in cognitive performance when age and education level are included, yet this decline is not as obvious in day-to-day tasks.Although some people with  will eventually develop dementia, others will remain in this  stage for a significant amount of time before progressing to dementia.Because of this,  is a disorder that can present very differently in different patients.In any case, research has been shown that patients diagnosed with  have a substantial risk of developing dementia by the third month following the onset of dementia symptoms.This risk was observed to increase significantly with time.The symptoms most commonly associated with  are those related to attention and executive function in .Daily functioning is unaffected by mild cognitive impairment.A decline in long-term memory, particularly episodic memory, is related to dementia, which is the next step following mild cognitive impairment.Mild cognitive impairment is the stage before dementia.10% of patients will acquire post-stroke dementia (PSD) or severe dementia in the months following the commencement of an ischemic stroke (30% with recurrent ischemic stroke).This can happen as early as three months after the stroke [3].
Electroencephalography (EEG) has been extensively employed in recent years to investigate the cortical abnormalities linked to dementia and cognitive decline [5].As a result, EEG signal analysis may reveal information on cognitive decline and dementia.Clinical EEG has a frequency range of 1 to 100 Hz and an amplitude of about 10-100 millivolts [6].The main differences between healthy people and those with a problem with working memory () can be summed up as follows: Dementia causes a slowing of EEG signals due to a power shift to lower frequencies and decreased corticalsubcortical communication for patients with  and  .Also, the reduction in signal complexity caused by dementia and other neurodegenerative diseases.Understanding the differences between how healthy people and people with  problems show EEG signals can help make clinical signs and better ways to diagnose neurological illnesses [5].
However, the EEG is impacted by extracranial sources known as artifacts, which can imitate the abnormal activity of the brain and hence impair the analysis.Such artifacts have been seen in EEG recordings and wrongly attributed to neurological disorders.Clinical studies involving EEG signals necessitate the creation of automatic algorithms to eliminate artifacts.Several methods have been proposed for artifact removal in the literature [7], including regression-based analysis, wavelet transform (WT), Independent Component Analysis (ICA), blind source separation (BSS), and epoch rejection.In another example of an automated hybrid artifact removal method, Al-Qazzaz et al. estimated ICs first, then used DWT to detect components as artifacts that had been marked and denoised.To produce an EEG devoid of artifacts, we correct the ICs and then reconstruct them using inv-ICA.The benefits of the proposed method were seen by the authors, who found that it improved discriminating between dementia and healthy groups [5,8].
Numerous studies have been conducted over the past few years to examine the impact of  and AD on EEG signals and how they change over time.Using resting-state EEG recordings, Yin et al. have devised a scheme based on integrated spectral and temporal analysis for the identification of  .Stationary wavelet transform (SWT) and descriptive statistical analysis with support vector machine (SVM) classifiers were used to establish a three-dimensional discrete feature space.A machine learning-based methodology has been presented by Kashefpoor et al. to distinguish between  and normal cases utilizing basic spectral frequency band EEG data [9].Eight EEG biomarkers, including power spectral density, skewness, kurtosis, spectral skewness, spectral kurtosis, spectral crest factor, spectral entropy, and fractal dimension, have been studied by Sharma [11].
Previous research has largely used a 2-way classification (  vs.  ,  vs.  ) however some studies have reported using a 3-way classification.Other researchers have looked into the possibility of using power spectrum analysis for the early diagnosis of AD and  [5].In [12] , neuro-markers based on complexity are calculated from AD patients and  participants.These complexity measures include fractal dimension and Lempel-Ziv complexity.A variety of entropies, such as spectral entropy ( ), permutation entropy (  ), Tsallis entropy (  ), approximation entropy (  ), and sample entropy ( ), could be computed for this demand in order to examine the EEG markers that may aid in dementia early detection [13].
A three-way categorization methodology was used in one study [14].Eyes-open, eyes-closed with a counting task, and eyes-closed conditions were used to compute spectral and complexity features from  ,  , and AD participants, respectively.Time-frequency domain parameters (relative power band, median frequency) and entropy-based neuro-markers (spectral entropy, sample entropy, auto-mutual information) have been merged by Ruiz-Gomez et al. [15].Eyes-open and eyes-closed EEG recordings were obtained from individuals with  ,  , and  [16].Classification was carried out using Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Multi-layer perceptron (MLP).Toural et al. have classified patients into three groups using wavelet entropy, relative beta, and theta power [17].In the preclassification phase, a SVM is employed for binary assessment, and in the classification phase, a neural network is implemented for the voting mechanism.The power spectrum density of EEG sub-bands and interhemispheric coherence were determined by Oltu et al. using data from , , and  .Each EEG sub-band's variance and amplitude sum, as well as the coherence amplitude sum, are included in the feature vectors [18].
The use of EEG for the detection and classification of dementia-related brain activity patterns has shown encouraging results.In spite of this, additional study is required in two areas: dimensionality reduction and channel selection.First of all, dimensionality reduction strategies try to minimize the loss of information by simplifying the number of features or variables in EEG data.Next-generation classification algorithms can be made more manageable in terms of both complexity and computing overhead if the dimensionality of the data is reduced.The best dimensionality reduction strategies to improve EEG categorization in dementia are not yet fully understood.Despite their success elsewhere, techniques for dimensionality reduction were applied to increase the classification accuracy.There are many techniques that can be applied, including the well-known principal component analysis (PCA) technique for dimensionality reduction.The PCA approach is frequently used to prevent redundancy in high-dimensional data [19].Additionally, channel selection is a type of feature selection that may be applied to the removal of irrelevant or noisy channels and the selection of channels with related features [20].The most efficient EEG channels have been determined using the sparse common spatial pattern (SCSP) algorithm, the mutual information technique [21,22], the recursive channel elimination (RCE) approach [23], and the differential evolutionbased channel selection algorithm (DEFS_Ch) [20,24].Although the strategy of channel selection can offer the benefits of eliminating unimportant channels or choosing a small number of significant EEG features to enhance classification performance.
Second, channel selection includes picking the EEG channels that add the most value to the categorization process.This method can enhance classification precision, however, the best way to pick channels for EEG classification purposes in dementia is not yet certain.Channel selection can be based on a variety of parameters, including statistics, spectral analysis, and geographical patterns; nevertheless, their efficacy and robustness must be assessed in the context of dementia [25].
Most of the approaches in the literature have a complicated structure and take a long time since they do not employ a data-efficient reduction technique, which is necessary for fast and accurate analysis.
Effective dimensionality reduction approaches and optimal channel selection procedures are the last pieces of the puzzle when it comes to improving EEG classification for dementia.Closing these knowledge gaps will aid in the creation of more precise and time-saving EEGbased categorization systems for the diagnosis and monitoring of dementia.To improve EEG categorization for dementia, more study is needed to analyze and compare different ways and determine the most suitable techniques.Based on background electroencephalography (EEG) activity, the goal of the current investigation is to differentiate between working memory () in five patients with vascular dementia (  ), fifteen post-stroke patients with mild cognitive impairment (  ), and fifteen healthy control individuals ().In the current study, as spectral entropy (), permutation entropy () and approximation entropy (  ), were the features that were selected to investigate the  and classify their tasks using the k-nearest neighbours (NN) classifier scheme.Therefore, a comparative study of using the fuzzy neighborhood preserving analysis with  -decomposition (  ) as a dimensionality reduction technique and the improved binary gravitation search (  ) optimization algorithm as a channel selection method has been conducted.The  was used in this investigation as a dimensionality reduction method to maximize the distance between the centers of various classes while minimizing the distance between samples that belong to the same class [26,27].Additionally, the most efficient channels that increase classification accuracy have been found using the  algorithm [25].Additionally, the suitability of  ,  , and  characteristics for the early identification of  was examined.NN classifier has also been used to identify patients with post-stroke  dysfunction.

Materials and Methods
To improve the  classification of dementia patients, the EEG signals would undergo various signal processing phases, as shown in Figure 1.The participants in this EEG study participated in a session of an auditory  task.The nonstationary EEG signals were initially processed using a wavelet (WT) denoising approach during the preprocessing step.After that, we look into and extract the meaningful features, such as non-linear , , and  entropy features, and conduct a comparison of the dimensionality reduction techniques fuzzy neighborhood preserving analysis with QR-decomposition () and the improved binary gravitation search (IBGSA) optimization algorithm for channel selection.Finally, the performance of the classifiers utilized is evaluated, showing that dementia classification techniques can be used to categorize patients' mental disability after stroke.

Participants
In the current investigation, 35 patients' EEG datasets were examined.The sample was recruited from the stroke unit and neurology clinic at the Pusat Perubatan Universiti Kebangsaan Malaysia (PPUKM).15  participants (7 male and 8 female,age 60.06±5.21),15  patients (5 male and 10 female,age 60.26±7.77),and 5  patients (3 male and 2 female,age 64.6±4.8) had their EEG data reviewed.The cognitive evaluations that were administered to the three groups were the mini-mental state examination (MMSE) [7] and the Montreal cognitive assessment (MoCA) [8]. participants' MMSE and MoCA scores were (29.6±0.73,29.06±0.88),but  patients' MMSE and MoCA scores were (20.2±5.63 and 16.13±5.97),respectively.Lastly, the MMSE and MoCA scores for the  patients were (14.8±1.92 and 14.8±1.92)respectively.The experimental methods utilized throughout the research were approved by PPUKM's Human Ethics Committee, and the patients' voluntary and informed consent was secured by acquiring signed consent forms.All patients were diagnosed using computed tomography/magnetic resonance imaging (CT/MRI) scans, patient medical histories, and clinical and laboratory tests.
A referential montage was created, and the channels were constructed according to the 10-20 international framework.The Nicolet EEG system was sampled at 256 Hz, and the electrode-skin impedance was tested to ensure that it did not exceed10 kilo ohms.This corresponded to a sensitivity of 100 v/cm, whereas the low cut-off frequency and high cut-off frequency were, respectively, 0.5 Hz and 70 Hz.The EEG was recorded for 60 seconds, with a 0.5-second fixation cue preceding the start of the recording period.The patients were then asked to commit five words to memory for 10 seconds as part of a simple auditory  test involving working memory.Following this, EEG recordings were made as each subject attempted to recall the phrase.After the 60-second interval elapsed, the researcher instructed the participants of the sample group to open their eyes and to recall in turn each of the words they had memorized [3].

Preprocessing Stage
In this investigation, WT denoising was utilized to eliminate EEG artifacts.Since the sampling frequency was 256 Hz [28], the symlets mother WT of order 9 'sym9' and 5 decomposition levels were utilized to decompose the acquired EEG information.
As in Equation 1, the discrete values of  and  can process the DWT.It can be constructed as  set of decomposition functions of the correlation between the signal () and the shifting and dilating of the mother wavelet function ().In Equation 2, location parameter  shifts MWT and the frequency scaling parameter  dilates or contracts it [29,30]

Features Extraction Stage
Each EEG dataset had 19 channels with a duration of 60 seconds, so 15360 samples were utilized for this study.Entropies have been utilized to identify anomalies in dementia patients' EEGs.The EEGs of dementia patients have been separated from those of age-matched healthy people using  [31][32][33].After normalizing the power spectral density (  ) to a scale from 0 to 1 to obtain normalized PSD (PSD n ), which has the value 1 for ∑ PSD n (f) = 1, the  is computed as in Equation 3 [31].
Utilizing the algorithm described in [34],  is calculated as in Equation 4 [35].(, , ) = ∅  () − ∅ +1 () … (4) where, ∅  is the natural logarithm for  contiguous observation within tolerance width  and  is the number of points of the EEG time series.For the sake of our analysis,  is calculated with a tolerance of r = 0.2 × SD and a run length of m = 2 epochs, where SD is the standard deviation.
In the case of , this sort of entropy is widely employed in the context of artefacts and noise , and one of its distinguishing properties [36] is its computational speed.
In terms of  's applications, non − stationary and non − linear signals are frequently employed [37]. has been utilized by researchers to assess the complexity of EEG signals in Alzheimer′s disease (AD) patients [38].The  can also be used to detect aberrant electrical activity in the brain, which cannot be demonstrated by traditional EEG signal detection methods [39].
When all motifs have equal probability, the largest value of  is obtained, which has a value of ln !, where  = 3,  = 1.In contrast, if there is only one (  ) different from zero, which illustrates a completely regular signal, the smallest value of  is obtained as much as 0 [36,40,41].For 60 seconds,  = 15360 samples, 6 windows of 10 second length (2560 samples) were extracted from the original EEG time series for each of 19 channels.
In order to estimate the  , assume the time series of  = { 1 ,  2 , … ,   } of length , at each time  of  a vector including the  ℎ subsequence value constructed as:   , = {  ,  +1 , … ,  +(−2) ,  +(−1) } for  = 1,2, … ,  − ( − 1) , where  is the embedded dimension, determines how much information is contained in each vector and  is the time delay.To calculate the  , the  of   are associated with numbers from 1 to  and arranged in increasing order as { +( 1 −1) ,  +( 2 −1) , … ,  +( −1 −1) ,  +(  −1) } for different samples, there will be !potential ordinal patterns,  , which are named "motifs" [38].For each   , (  ) demonstrate the relative frequency as follows: Where #{ } denotes the cardinality of the set (the number of elements).The  is computed as follows: ln (  ) …( 6) When all motifs have equal probability, the largest value of  is obtained, which has a value of ln !, where  = 3,  = 1.In contrast, if there is only one (  ) different from zero, which illustrates a completely regular signal, the smallest value of  is obtained as much as 0 [36,40,41].For 60 seconds,  = 15360 samples, 6 windows of 10 second length (2560 samples) were extracted from the original EEG time series for each 19 channels.

Statistical analysis
The denoised 19 channels from the EEG dataset of 15  , 15  , and 5  patients were preliminary divided into 5 recording regions that correspond to the scalp area of the cerebral cortex.The frontal includes seven channels: Fp1, Fp2, F3, F4, F7, F8 and Fz, temporal includes four channels: T3, T4, T5 and T6, parietal includes three channels: P3, P4 and Pz), occipital includes two channels: O1 and O2), and central includes three channels: C3, C4 and Cz).The Kolmogorov-Smirnov test determined normality, while Levene's test confirmed homoscedasticity.Thus, SPSS 22 used two analysis of variance (ANOVA) sections on , , and  characteristics.Each segment had two independent variables (IVs): the subject groups (  subjects,  , and  patients) and the five scalp regions (frontal, temporal, parietal, occipital, and central).One of the former attributes was the dependent variable (DV).All statistical tests were significant at P< 0.05.

Dimensionality reduction using 𝐅𝐍𝐏𝐀𝐐𝐑
In order to maximize the distance between the centers of various classes while minimizing the distance between samples that belong to the same class, this study also used the fuzzy neighborhood preserving analysis with QR-decomposition (  ) dimensionality reduction technique [26] of Khushaba et al [20]. maintains the contribution of samples to various classes in this way [26].For the first time, our study used  to distinguish between  and demented participants during  tasks.The matrix (  ) was built from the training set to project the input feature vector using  .To reduce dimensionality, the projection matrix was multiplied by the training and testing sets. projected training data input feature vector.Projecting the feature vector's testing set requires merely multiplying it by the projection matrix from the training data.Figure 2 shows how the  feature projection calculates the within-class scatter matrix (  ) and between-class scatter matrix (  ).
calculates the within-class scatter matrix (   ) and between-class scatter matrix (  ).

Fig. 2. The steps of Dimensionality reduction using 𝐅𝐍𝐏𝐀𝐐𝐑
In a comparative study, the  dimensionality reduction approach and  NN classifier were used to identify  ,  , and  subjects [5].

Channel Selection using 𝐈𝐁𝐆𝐒𝐀
The most effective channels have been found, and the amount of information has been decreased, using the improved binary gravitation search algorithm (  ) optimization algorithm [25,42]. is a powerful optimization technique that was first proposed in [43].for use in addressing binary-valued problems.It was created based on the Newtonian laws of gravity and motion. objects (agents) are defined for the  algorithm to determine the best EEG channels.The population starts out with this collection of things.Each object in this study is regarded as a binary vector with a dimension of 19.The number of EEG channels is the same as the dimension that was given.The following vector can be regarded as the  ℎ object.Finding the item that gives the best fitness value is the main objective.Equation 7 [44] can be used to calculate the classification accuracies for each set of EEG channels, which are used in this work to determine the  values for the objects: . . .where  1 ,  2 are predefined weight factors,  1 is the weight factor for the classification accuracy of the  NN classifiers respectively determined by the 10-fold cross-validation () method;   is the 1-NN classification accuracy;  2 is the weight factor for the number of selected features and   is the value of the feature mask.If precision is the most crucial factor, the ℎ factor might be increased to a high amount (such as 100%).The position of the object with a high fitness value should be set suitably since it has a high likelihood of influencing the positions of the other objects in the following iteration [14,15].Equation 8 yields the   , where corr is the number of cases that were properly classified and incorr is the number of examples that were classified wrongly [44]: The IBGSA algorithm selects the most informative EEG channels for classification.The method selects the optimal EEG channel subset for classification.
The technique optimizes channel selection to increase EEG-based classification accuracy and efficiency [43].

Dementia Classification Techniques
The EEG signals were divided into (, , and ) using a NN classifier.The patients with  made up a statistically significant minority in this analysis.To address the discrepancy, the researchers used a synthetic oversampling technique (SMOTE) [45].To prevent overfitting and bias in the classification analysis, the classifier parameters and the percentage of oversampling were evaluated by 10-fold cross-validation with a grid search approach [46].The provided data set was partitioned into ten independent samples of similar size.Only one of these groups was utilized to train the classifier, while the other nine were used as the test set.Ten iterations of this process yielded ten reliable results.The 10-fold CV accuracy of this dataset was calculated as the mean of these accuracies [47].
Since SMOTE modifies the dataset, the oversampling was incorporated into the settings.Because of this, it is possible that the parameters discovered with varying amounts of SMOTE are not equivalent.When using the SMOTE to normalize the class frequencies, we solely considered the training set [48,49].The classifier in this study was trained to find the best value of k, which was discovered to be k = 5, and to increase classification accuracy.Each trial has been classified by NN using the Euclidean distance as a similarity metric.
After selecting the optimal EEG channels, the NN algorithm classifies background signals from symptomatic and asymptomatic instances using entropy. NN uses labeled training samples to classify data points.EEG signal entropy measures randomness or chaos.The NN method classifies symptomatic and asymptomatic cases based on their resemblance to training samples by computing the entropy of chosen EEG channels [50].

Results of Statistical analysis
Patients with  showed less complexity than those with I and  using ,  and  as shown in Figures 4, 5 and 6.For all patients, but especially for those with  and , the complexity of the EEG signals reduces as the condition worsens.
Figure 4 shows that the  values of the  patients were lower than those of the  patients and that the  patients' values were the greatest.
Moreover, the  patients had lower  values than the  patients, and the  subjects had the highest  values (Figure 5).Finally, the patients with  had lower  values than those with , and the participants with  had the greatest value (Figure 6).

Results of Dementia Classification Techniques
The classification confusion matrix for all three schemes proposed is shown in Tables I, II and III.As a result,  NN was used in the study to support multi-class classification and to distinguish ,  patients, and  subjects.This study had several limitations, including a small sample size, and an additional analysis with a large database should be performed in the future.
Finally, we compared the proposed approach to other cutting-edge methods in the literature that employed the different dementia datasets as we did in our work.The results show that our model outperforms other existing methods in the literature, with the highest classification accuracy of 90.52% compared to Kashefpoor et al. [9] proposed a methodology that obtained an accuracy of around 88%, Sharma et al. [10] investigated different features for control vs.  signal classification and obtained accuracy ranges between 73.2% and 89.8%.

Conclusion
The electroencephalogram (EEG) is a vital tool for studying mental processes.Here, we analyze and filter EEG signals to identify promising channels and useful markers for an earlier, more accurate diagnosis of dementia.The WT method has been implemented as a denoising method.Patients with  and  have had their irregularities assessed with , , and  as characteristics.To improve  categorization, we applied the  dimensionality reduction technique in conjunction with the IBGSA channel selection algorithm. dimensionality reduction technique increased  NN classification accuracy from 86.67% to 88.09%, while the IBGSA channel selection algorithm increased it to 90.52%.Findings revealed that IBGSA reliably enhances  discrimination in  ,  patients, and  controls.Because of these findings, it is clear that minimizing the number of channels used in the IBGSA selection process has a substantial impact on improving classification accuracy.
Working memory stores and manipulates information for ongoing tasks.Dementia causes cognitive deterioration, including working memory loss.Therefore, working memory enhancement may slow dementia-related cognitive deterioration.Working memory training may increase brain activity in working memory-related areas, however, healthy volunteers trained working memory using an adaptive N-back task.The training did not directly cause changes in brain activation in several important locations.In spite of that, studies have implied that working memory training may have benefits, but more study is needed to determine the long-term consequences.Working memory training alone may not diminish dementia-related brain activity since dementia is complicated.Working memory training may also be affected by dementia stage, type, and training procedure.In general, improving working memory in people who have dementia continues to be our focus of research.Future studies may provide additional insights into effective interventions and techniques to improve cognitive function and quality of life.

Fig. 1 .
Fig. 1.The block diagram of the proposed study.

Figure 3 Fig. 3 .
Figure 3 depicts the denoised EEG signals produced by the WT method.Due to the heterogeneity of EEG artifacts, WT has been

Fig. 4 .
Fig. 4. Comparative plot of the  for the five scalp regions of the brain for ,  patients and  subjects.

Fig. 5 .
Fig. 5. Comparative plot of the  for the five scalp regions of the brain for ,  patients and  subjects.

Fig. 6 .
Fig. 6.Comparative plot of the  for the five scalp regions of the brain for ,  patients and  subjects.
et al. for the diagnosis of individuals with  [10].
Table I illustrates the confusion matrix for the NN Classifier without using the  and IBGSA.Table II shows the confusion matrix for the  NN Classifier with the  dimensionality reduction technique.The number of selected features by the  dimensionality reduction technique was set to 40 characteristics, which are the most essential features in terms of differentiating between patients suffering from  and stroke-related  and healthy control participants.Table III shows the confusion matrix for the NN Classifier with the IBGSA channels selection algorithm.The  NN classification accuracy was improved from 86.67% to 88.09 using  dimensionality reduction technique and 90.52 by IBGSA channel selection algorithm.The results suggested that IBGSA consistently improves  discrimination of ,  patients and  subjects.Therefore, IBGSA improves the classification over all accuracy for all three groups as in the  NN classification over the  accuracy.