IEEE SENSORS JOURNAL, VOL. 21, NO.

13, JULY 1, 2021 14263

Statistical Wavelets With Harmony Search-

Based Optimal Feature Selection
of EEG Signals for Motor
Imagery Classification
Samrudhi Mohdiwale , Mridu Sahu, G. R. Sinha , Senior Member, IEEE,
and Vikrant Bhateja , Senior Member, IEEE

Abstract —Brain Computer Interface (BCI) does not only

help people of physical disability but also being used pop-
ularly in several other applications. Motor Imagery (MI) classi-
fication is one of the major contributions in BCI which works
on segment of EEG signal within particular frequency band.
Herein, feature selection plays important role in obtaining
good classification results. In this article, Beta and Gamma
frequency are considered with Statistical DWT (SDWT) based
features for classification of EEG signals (MI classification) for
patient monitoring, assistance healthcare services and daily
living activities. Harmony search algorithm of feature selec-
tion is used to obtain the optimal feature set for classification
of MI. The results show that frequency centric SDWT achieves
average accuracy of 92.49% for weighted KNN (K-Nearest Neighbour) method. Comparison of accuracies before and
after feature selection portrays that feature selection with harmony search improves the performance of proposed MI
classification.
Index Terms — Brain computer interface, EEG, motor imagery, wavelet transform, optimal feature selection, harmony
search optimization.

I. I NTRODUCTION brain and used in several applications [1]–[3]. It is interesting

to note that in motor imagery (MI) task classification, no exter-
T HE electroencephalograph (EEG) is the most prominent,
non-invasive and less costly signal acquisition technique
which is used in brain computer interface (BCI) for com-
nal stimulus is required to perform imagination task, which
is the most important and sensitive part of other BCI based
munication between brain and machine. Communication for applications such as mental arithmetic and speller designs.
imagined motor movements such as hand or feet movement In motor imagery task classification for BCI, EEG signals are
from brain to system is termed as motor imagery task for BCI mostly used due to their non-invasive nature of acquisition
which is useful in assistive robot technologies controlled by which results in no cortical damage in brain than other invasive
technique but suffers from high artifacts. These artifacts result
Manuscript received August 10, 2020; accepted September 8, 2020. in reduced accuracy for BCI performance. The EEG signals
Date of publication September 23, 2020; date of current version June 30, are non-stationary in nature and become a challenging task
2021. The associate editor coordinating the review of this article and
approving it for publication was Dr. Ing. Emiliano Schena. (Corresponding to work with these signals and to mitigate this, the signals
author: Samrudhi Mohdiwale.) are segmented into small time frames in which the signal is
Samrudhi Mohdiwale and Mridu Sahu are with the Depart- assumed to be stationary. Various studies have been reported
ment of Information Technology, National Institute of Technology at
Raipur, Raipur 492010, India (e-mail: [email protected]; for MI task classification [4], [5] etc. using time domain,
[email protected]). frequency domain and common spatial pattern as features
G. R. Sinha is with the Myanmar Institute of Information Technology, for motor imagery classification task. In spite of numerous
Mandalay 05053, Myanmar (e-mail: [email protected]).
Vikrant Bhateja is with the Department of Electronics and Commu- techniques on motor imagery classification reported in current
nication Engineering, Shri Ramswaroop Memorial Group of Profes- literature, there remains a potential scope in the area best
sional Colleges (SRMGPC), Lucknow 226028, India, and also with the suitable technique that does not only classify the MI tasks but
Dr. A. P. J. Abdul Kalam Technical University, Lucknow 226031, India
(e-mail: [email protected]). also reduces the complexity of the model and the feature space.
Digital Object Identifier 10.1109/JSEN.2020.3026172 It is important to extract the relevant features simultaneously
1558-1748 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 14264 IEEE SENSORS JOURNAL, VOL. 21, NO. 13, JULY 1, 2021

to reduce the irrelevancy of features along with feature space. Some recent techniques on MI task classification include
To work on this, we have studied different features and feature multiplex weighted visibility graph [17]; energy principal
selection techniques which are applied in various applications. component analysis based feature extraction and deep belief
Based on the study, a method is proposed to effectively classify network [18]; and improved channel selection for CSP MI
the MI tasks. The details are provided further in the paper. task classification [19]. However, these techniques have short-
The paper is organized as follows: Section 2 prensents comings that either it required subject wise calibration or uses
related work; In section 3, Motivation and Problem Formula- deep learning technique which improves the overall results on
tion are discussed; Section 4 describes the datasets used along the expense of complexity of the model. The small training
with proposed method of sub-band based statistical wavelet for samples generates the need of effective and suitable data
feature extraction and harmony search based feature selection augmentation technique.
method. Section 5 discusses the results of proposed method
with analysis of the results and comparison with other existing B. Related Work for Feature Selection and Optimization
method. Finally, the section 6 concludes the paper. Techniques
Feature selection is one of the important aspect that is useful
to reduce the feature space and complexity as well as to get
II. R ELATED W ORK the relevant features. Based on taxonomy of feature selection,
evaluation strategy of feature selection can be categorized
A. Related Work for MI Feature Extraction
into 5 major categories i.e. Filter method, Wrappers method,
Common spatial pattern (CSP) is widely used technique for Embedded method, Hybrid method, Ensemble method [20].
Motor imagery task classification which works on segment Filter method is dependent on selection criteria such as infor-
of EEG signal with particular frequency band. The concept mation, distance, dependency etc however, it is independent of
of CSP is referred as optimization problem for maximizing learning algorithm. Minimum redundancy maximum relevancy
the variance of positive class and simultaneously minimizing (mRMR) [21], entropy, information gain [22], Relief-F [23],
the variance of negative class [6]. The CSP features focus Chi-Square [24] are some of the techniques of filter method.
on spatial property of the signal not on the spectral one and This technique is used in the channel selection and feature
therefore choice of frequency bands and specific frequencies selection problems for different EEG based applications such
becomes important issue in common spatial patterns [7], [8]. as emotion recognition [25], stress detection [26] and many
The CSP also suffers from problem of overfitting because it is more. Wrappers method uses the dependency among the
based on L-2 norm method [6], [9] that can be overcome by features and it is classifier specific. Different combinations
Regularized CSP and L-1 norm based CSP’s [10], [11]. These of features are tested on the model and provide the best
CSP methods are computationally intensive. combination based on the forward or backward selection [20].
Considering the above CSP aspects and limitations of Optimization method can also be used with wrappers method
various methods, frequency or time frequency characteristics to provide the optimal subset of features. Feature selection
of motor imagery EEG signal are mater of concerns. While with wrapper method is used in various literature for EEG
performing MI task, event related synchronization and desyn- tasks categorization [27], [28].
chronization take place in sensory motor rhythms mainly in Hybrid and Ensemble methods are also popular in few
[μ, β] and γ frequency band respectively [5], [12]. Power applications. When optimization based methods are considered
Spectral Density (PSD) is also highly explored feature in MI for MI tasks, to improve the diversity of ensemble classifier,
task categorization. sub-band based CSP with Fuzzy integration of PSO is sug-
Further, we investigated literature for support on frequency gested which introduced information of uncertainty, incom-
based features or their extension in addition to CSP based pleteness and imprecision of EEG, generally skipped during
methods. Park et al. [13] proposed frequency based multi- feature extraction, via fuzzy integration method. Fuzzy method
variate empirical mode decomposition to localize time fre- PSO is used for improved confidence [29]. Another study
quency representation. Kevric and Subasi [14] demonstrated of optimization based MI task classification uses Differential
the comparative analysis of decomposition based methods and Evolution method for feature selection. Wrappers method of
found that higher order statistics features are more prominent feature selection best suits to check the performance of opti-
over other features extracted from DWT, WPD with sub- mization algorithm [30] Harmony search optimization is one
bands. It also provides comparative analysis of DWT and of the popular optimization technique nowadays due to its lim-
WPD based method with CSP based approach and found that ited mathematical requirement and successful due to its insen-
wavelet based approach more consistent in terms of perfor- sitiveness to initial value. The method proved its efficiency
mance. Mirvaziri and Mobarakeh [15] adopted DWT feature in various machine learning problems and shows various
extraction technique and attempted to improve the classifier advantages over existing traditional optimization methods [31]
with PSO based algorithm. Luo et al. [16] investigated the hence, this technique is further utilized in the current work.
effect of ERS/ERD and spatio-temporal discrepancy feature
is suggested for improvement in signal to noise ratio (SNR). III. M OTIVATION AND P ROBLEM F ORMULATION
Welch power spectral density, Burg’s Auto-regressive feature, A comprehesive study on MI task classification using dif-
and Wavelet packet decomposition are extracted and classified ferent methods has been discussed so far. The study suggests
with Support Vector Machine. the need of a model that can effectively classify MI task for

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 MOHDIWALE et al.: STATISTICAL WAVELETS WITH HARMONY SEARCH-BASED OPTIMAL FEATURE SELECTION 14265

different applications such as patient monitoring, home and

healthcare assistance. Advantages of frequency based features
over CSP features enhances the interest towards frequency and
time - frequency based feature extraction technique. Popularity
of optimization techniques due to successful implications in
different fields motivate to use the technique for MI task
classification for performance improvement.
Keeping this in notion, the paper proposed sub-band selec-
tion based statistical wavelet features for MI task classification.
Sub-bands referred to β and γ band represent the character-
istics of MI task [32]. We aim to: (1) Improve classification
accuracy for MI task and (2) Analysis of optimization tech-
nique based feature selection for MI tasks. Wrappers method
of feature selection with harmony search optimization has been Fig. 1. Block diagram of proposed methodology.
employed for task characterization. Moreover, the features
obtained from selection technique are best suitable, further
passband frequency of 0.5Hz and Stopband frequency 100Hz.
utilized in different classifiers to improve performance. Finally,
Maximally flat response and uniform passband property of this
comparison between proposed method and different existing
filter makes it more suitable at the cost of wide transition band.
methods is done based on accuracy.
The Butterworth filter is mathematically expressed as (1).
1
IV. M ATERIALS AND M ETHOD H (x, y) = 1 − (1)
D(x,y)W 2n
This section discusses dataset used and proposed method- 1 + [ D(x,y) 2−D 2 ]
0
ology for MI task classification which includes description of
where, D(x,y) is the distance from center frequency of window
dataset, methods for feature extraction, feature selection and
used from points x, y; W is the bandwidth of filtering range;
classification.
and n is the order of the filter. Choice of order of the filter
depends upon the tolerance of transition band as well as
A. Dataset number of datapoints available per sampling frequency. After
The current study is based on open source datasets available preprocessing operation, the data is segmented according to
for MI task classification. The dataset is taken from BNCI the values provided in the description of dataset for each task
Horizon 2020 website which dataset consists of 10 subjects. for both the dataset.
Cue guided grazed paradigm was used for recording. The 2) Feature Extraction: Signal obtained from segmentation
task performed in a single session contains 8 runs (5-Training procedure has high dimension and also holds different rele-
and 3-Validation), each run having 20 trials. The subjects vant and irrelevant information. Feature extraction procedure
were asked to perform kinesthetic MI of right hand and feet, provides systematic approach to extract relevant information
sustained for 5 sec. The recordings are taken from Ag/AgCl and simultaneously discard irrelevant one. To choose the rel-
electrodes placed on C3, C4, Cz positions and neighbouring evant characteristics, time frequency decomposition technique
positions on the central lobe. Detail description of dataset is is utilized as widely used method. The Discrete Wavelet
discussed for further understanding in [33], [34]. Transform (DWT) is chosen for MI task categorization based
on the different literature which suggests that DWT is suitable
technique for non-stationary signal such as EEG. This tech-
B. Proposed Method nique localizes the changes in frequency of signal over time
The workflow of proposed method is shown in Fig. 1 which and thus both time and frequency information is taken into
highlights the overview of MI task classification. The method- consideration unlike CSP features with reduced computational
ology consists of three steps: initial step is data composition in complexity. However, Short Time Fourier Transform is also
which data is prepared as per the need of experimental design, similar but doesn’t provide amplitude fluctuation information
second step include details of feature extraction method and and reduced resolution so DWT becomes popular measure
in final step, feature selection classification and performance [35], [36]. In the current work, the signal is decomposed
evaluation are performed. Each block of proposed method is into 5 level wavelets with mother wavelet as ‘dB 4’. As the
discussed in detail in further subsections. name suggests that the wavelet has (N/2= 4/2) 2 vanishing
1) Data Decomposition: The experimental data has been moments with 5 detailed coefficients and one approximate
imported into the workspace for further processing. coefficient. The DWT of level 5 is chosen as per the frequency
Data taken from electrodes is unprocessed and contain range of interest, 0-100 Hz. See Table I for EEG sub-band
various artifacts; one such artifact is human function arti- frequency range in 5 level DWT decomposition.
fact which can be reduced during preprocessing stage from From the literature, γ and β waves are responsible for event
the original signal. In this stage, undesired frequencies and related synchronization and desynchronization of motor neu-
information are removed. To preprocess the data, 5th order rons in EEG signal hence, these are taken for further process-
Butterworth bandpass filter is employed which is having ing in feature extraction of EEG signal. Statistical features of

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 14266 IEEE SENSORS JOURNAL, VOL. 21, NO. 13, JULY 1, 2021

TABLE I TABLE II
F REQUENCY R ANGE OF D ETAILED C OEFFICIENTS IN 5 L EVEL DWT S IMULATION PARAMETERS FOR H ARMONY S EARCH O PTIMIZATION

wavelets are powerful technique of feature extraction. Differ- classification accuracy. The fitness is calculated as:
ent statistical features such as standard basic statistics (mean, Fi tness = Maxi mi ze[ Accur acy(Fk )] (2)
median, standard deviation etc.) and higher order statistics
(entropy, kurtosis, skewness etc.) are explored for EEG signal Here the fitness is function of Features (Fk ) which are also
processing for different applications [14], [37] [38]. It pro- known as decision variable. Therefore, selection of features
vides better results than only coefficient based features and plays an important role. Initially to select the features, binary
hence statistical features of wavelet coefficients of chosen vectors are generated randomly having the size similar to
sub bands with energy are calculated. The statistical features Feature space. In this method, musicians are decision variables
used in the study include mean, standard deviation, variance, for the fitness function; Notes represent values of the deci-
maximum value of detailed coefficient along with power of sion variable; Harmonies are number of populations; fitness
sub-band. Total 10 features are extracted in the study for each function is the cost function that needs to be minimized
trial. Feature vector for each trail is arranged as: of maximized; and Harmony memory represents the set of
good solutions. The algorithm also uses parameters like Har-
mony memory consideration rate and Pitch Adjustment Rate.
[Mean β , stdβ , V arβ , Max β , Powβ ,
These parameters determine the selection or widening the
. . . . . . . . . .Mean γ , stdγ , V arγ , Max γ , Powγ ](1×10) search space and selection of neighbouring values. The Algo-
rithm 1 represents the stepwise evaluation of feature selection
These features are further used in feature selection method method. Input to the algorithm includes feature vector, class
to get optimal set of features. labels, simulation parameters, max iterations and count of the
3) Feature Selection Using Optimization: The idea of feature features. The algorithm provides the optimal feature subset
selection becomes important to obtain maximum informa- that will provide maximum classification accuracy as per
tion from minimum features so as to reduce complexity the objective function. Initially, fitness function is defined
and improve efficiency of classification. Wrappers method of and classifier output in the form of accuracy is obtained
feature selection is most feasible for classifier based feature from Weighted KNN classifier. Based on count of number
selection problem. The current work also uses Wrappers of features, initial harmony positions are generated. Cost, i.e.
method of feature selection. However, Wrapper method is fitness value, initialized as empty matrix. In the first for loop,
used for feature selection but, the selected features are further populations are initialized randomly and for each set cost have
utilized in different classifiers to improve MI task classification been calculated. Best cost and best solutions are preserved for
accuracy [39]. evaluation in next iterations. Harmony search optimization is
Harmony search (HS) based feature selection technique is started, for the selected maximum number of iterations, new
used in the current work due to its advantages over other harmony values are generated and these are compared with
traditional techniques. This is population based approach in HMCR. Since HMCR is higher (near to 1), new solutions will
which, the new harmony encodes the feature subset based be obtained from existing HM, otherwise it will be generated
on the selection of recent best feature stored in the harmony randomly. The PAR value decides the mutation similar to
memory. Moreover, it can identify multiple best feature subsets Genetic Algorithm however HS algorithm uses members of
than the single one which reduces the chance to get stuck in HM for this purpose. After the whole procedure, new selected
local minima [40]. features are made available and the fitness is evaluated and
The method is inspired by improvisation process of music saved in fitness vector. The whole process is continued until
players. The stochastic nature of the algorithm provides the desired iteration has been achieved and stopping criteria
more exploration in the search space and less chances to are met. The output of the algorithm provides a set of optimal
stuck in the local minima. This property makes the harmony features with maximum accuracy.
search method more advantageous than the traditional method. Simulation Parameters: Simulation parameters used for
Another advantage of the method is that, it is non-sensitive to feature selection in harmony search optimization are given
initial values. Keeping this in mind, harmony search method in II. Input arguments are listed in which number of features
is utilized for feature selection in the current work. In any are, extracted features from MI dataset; Class represents cate-
optimization method, two main components are objective gorization problem (Hand/Feet movement); Harmony Memory
functions and rules that update the solution. In the current Size is the limit up to which harmony can store the previous
research, the aim is to use optimization to maximize the played notes; Harmony Memory Consideration rate is the ran-

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 MOHDIWALE et al.: STATISTICAL WAVELETS WITH HARMONY SEARCH-BASED OPTIMAL FEATURE SELECTION 14267

domness on choice of feature selection; PAR is the threshold TABLE III

value for updating feature and number of iteration is the max PARAMETERS U SED IN T REE C LASSIFIER
limit of run to update the best solution.
The parameters such as HMS, HMCR and PAR are selected
based on the experiment. The various combinations are tested
for parameter tuning. The selected parameter results in opti-
mal feature selection that maximizes the accuracy. Then, the
change in parameter may also increase the accuracy but at
the cost of higher computation or higher complexity. The
presented combination provides trade-off between accuracy
and complexity. The objective function converges for given MI task classification is considered. From the literature, KNN
set of parameters. and Tree based classifiers are best among other classifiers
hence, these classifiers with their variants have been chosen.
Algorithm for Feature Selection Using Harmony Search The 10 dimensional feature vector is obtained for each subject
Input: FeatureVector, Class, PAR, HMS, HMCR, MaxIt, used for classification of hand and feet moment hence, termed
Number of features(n) A = {a1 , a2 , . . . , a10 } of as binary class classification problem. Then 5-fold cross vali-
integers dation is employed for classification.
Output: Best feature subset, Best cost • Tree Classifier
FitnessFunction ← classification accuracy from Weighted Tree based classification is non parametric approach
KNN heuristic search technique for better classification. How-
Harmony Position (Pos) = []; ever, the method has lower computational complexity
Cost = []; and has capability to solve overfitting problem. Irrelevant
for i ← 1 to HMS do attribute have less impact on classification accuracy for
Pos = r and(0, 1)n this classifier. Choice of number of splits make the
SelectedFeaturePos ← (Pos == 1) tree simple to complex. Some ensemble methods also
FeatureSet ← FeatureVector (:,SelectedFeaturePos) provide enough coverage of features for better classifi-
Cost = FitnessFunction (FeatureSet, Class) cation. Complex tree has maximum splits as 100. As the
end decision tree is prone to variance and bias so, bagged and
HM ← Sort(Cost) subspace tree based methods are used. Bagged trees are
Best Sol ← HM(First Value) ensemble classification tree which select the number of
for itr ← 1 to MaxIt do features from split criteria and implement random forest
NewHarmony ← empty Harmony of size n algorithm (generally) on each split, combine the solution
for k =1 do to provide classification result [41]. Parameters used in
NewHarmony.Pos ← random number of size n tree classifiers are shown in III. The parameters are the
for j ← 1 to n do standard parameters presented in the classifiers.
if GeneratedRandomNumber < HMCR then • KNN Classifier
r ← RandomNumber of size 1 to HMS KNN is proximity based lazy classifier which works
NewHarmony(k).Pos(j) ← HM(r).Pos(j) on instance-based learning method. Instead of building
else if GeneratedRandomNumber < PAR global model best fitted for entire training data, it works
then on local model and local information-based classification.
NewHarmony(k).Pos(j) ← 1-HM(k).Pos(j) This might be susceptible to noise but provides flexi-
end ble decision boundary unlike tree and other rule-based
end classifiers. Although KNN provides flexible decision
end boundaries, it very prone to features of interest if not
SelectedFeaturePos ← (NewHarmony(k).Pos(j) == normalized. Features should be in some compatible range
1) to obtain better classification. KNN have many variants
FeatureSet ← FeatureVector (:,SelectedFeaturePos) based on values of K and distance measurement approach
Cost = FitnessFunction (FeatureSet, Class) between instances. Majority voting approach is adopted
end to predict the class of test data:
HM ← Merge HM and NewHarmony
SortOrder ← Sort(Cost) Ma j ori t yV oti ngC = argmax v K (v = Ci ) (3)
HM ← HM(SortOrder)
where Ci is the class, v is the label from one of the
Truncate HM upto given HMS
nearest neighbor, function K(.) returns the binary value
BestSol = HM(Intial Value) BestCost ← BestSol
of 1 if true else 0 [34]. In the current study, Fine
end
KNN, Weighted KNN, Subspace KNN are adopted for
investigation. These KNN variants are chosen due to
4) Classification: The classification operation is performed their optimal K nearest neighbors’ value with change in
over extracted and selected features. To select the best classi- distance measurement technique suitable for the dataset
fier, well known classification techniques which best suits for which rejects the outliers efficiently. Table IV shows
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 14268 IEEE SENSORS JOURNAL, VOL. 21, NO. 13, JULY 1, 2021

TABLE IV TABLE VI
PARAMTETERS U SED IN KNN C LASSIFIER P RECISION , R ECALL AND F-S CORE U SING W EIGHTED
KNN C LASSIFIER

TABLE V
R ESULTS OF β AND γ DWT F EATURES FOR MI TASK C LASSIFICATION
feature with dimension 1260 × 10 for each subject. The
experiment was performed for 5- fold cross validation of each
subject.
The results shown in V represents comparative analysis
of variants of Tree and KNN classifier. From the Table V
weighted KNN shows better performance among all the classi-
fiers for approximately all the subjects. The better performance
of Weighted KNN is further utilized in feature selection
method.
The precision recall and F Score are also reported for
Weighted KNN classifier. The results presented in Table VI
are evaluated with this classifier because it shows lesser
complexity, higher performance and doesn’t introduce bias in
the algorithm. Which is further utilized in the paper.
Parameters used in the KNN classification. Which are
From the results, Area Under the Curve (AUC) is shown in
the standard parameters presented in the classifiers. The
Table V, AUC for weighted KNN and bagged tree both are
results are shown in the subsequent sections.
0.98, which shows the reliability for further use.

V. R ESULTS AND D ISCUSSION B. Performance of Harmony Search Based Feature

Selection Method
The experiment was performed on MATLAB software with
system configuration intel core i5 processor and 8 GB RAM. The another objective of the paper was to evaluate the
Open source dataset is tested for offline analysis of the performance of optimization technique for feature selection in
proposed scheme of motor imagery task classification. MI task classification. Harmony search optimization is used
The experiment was performed on motor imagery task in the current work for feature selection due to the advantages
classification dataset as presented in section IV-A. To achieve presented in IV-B.3. The experiment was performed 25 times
the promises set in section III, different experiments have been for each subject and average accuracy has reported. In the
performed. For improved classification accuracy, statistical experiment, feature selection is applied to whole dataset as
wavelets of beta and gamma frequency band are extracted we are using 5 fold cross validation technique. This ensure
and analyzed, the obtained results are presented further in this there would be no bias in the experiment due to lack of rep-
section. The later part of this section represents the result resentation of samples in training and testing set. The average
obtained from harmony search optimization based feature accuracy from all the 5 fold data evaluation has reported.
selection method. Comparative analysis of proposed appraoch Fig. 2 shows the average accuracy obtained after 25 runs. The
with existing appraoches also presented in this section. The Fig. 2 also compare accuracy before and after optimization for
results used for the comparision in VII, VIII are directly taken Weighted KNN classifier. It can be said that feature selection
from the research papers. The significance of proposed method using harmony search method shows significant impact on
also analyzed with statistical test presented in V-D. classification accuracy. Feature selection also reduces the
feature space upto 6 features from 10 features.
Frequency of features selected in 25 runs are shows that the
A. Performance of Beta- Gamma sub-Band Based higher the occurance the more the importance, but it can be
Features seen that, no single feature is dominating. The combination of
This experiment was performed on MI dataset discussed features shows more impact on classification task.
in section IV-A, beta and gamma frequency based statistical
DWT (SDTW) technique of feature extraction is employed. C. Comparative Analysis
The experiment have been performed on all 10 available This section deals with comparative analysis of proposed
subjects in which left hand and feet movement have to be algorithm with existing algorithms. Common spatial patterns
classified. The resultant feature matrix is obtained after SDWT are widely used features for MI task classification however,

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 MOHDIWALE et al.: STATISTICAL WAVELETS WITH HARMONY SEARCH-BASED OPTIMAL FEATURE SELECTION 14269

TABLE VIII
P ERFORMACE C OMPARISON OF P ROPOSED A PPROACH
W ITH DWT B ASED F EATURES

Fig. 2. Impact of feature selection in MI task classification.

TABLE VII
P ERFORMANCE C OMPARISON OF P ROPOSED A PPROACH
W ITH CSP B ASED M ETHOD
Fig. 3. Results of statistical test.

D. Statistical Analysis
Statistical analysis is performed for the different algorithms
used for classification of motor imagery task. The Friedman
test is employed for statistical comparison on proposed method
with other existing methods because this test allow to compare
multiple algorithms [45]. Null hypothesis suggests that the
methods are same with alternative hypothesis is that the
the results for given dataset using CSP method have alot of methods are significantly different. Nemenyi post hoc analysis
scope of improvement. The results of proposed appraoch are is employed to test the best method among available ones [46].
compared with CSP method as presented in table VII. It can be The results from Friedman test shows the “p” value 0.0029
seen that the propsed appraoch shows higher performance than which is less than significance level 0.05. since p< 0.05, null
the CSP based method. In table VIII Wavelet based features hypothesis can be rejected and can be concluded that methods
are compared with the proposed algorithm. The work reported are statistically different. Further, Nemenyi post hoc analysis
in comparision used the wavelet coefficients as feature and results are presented in Fig. 3. The CD is the confidence
perform classification and another literature reported results value obtained from Nemenyi post hoc analysis. The Fig. 3
using PCA based feature compression however we are using represents the ranking of methods. The best method has rank 1,
statistics of wavelets which allows to represent the series of next best get rank 2 and so on. From the figure it can be
features as a single value. seen that the proposed feature selection approach gets 1st rank
There is large difference in the results of proposed algo- among all methods. The proposed approach with β and γ
rithm and compared algorithms due to the fact that, for the features also gets 2nd rank which shows the significance of
table VII, extracted features are CSP features, which suffer the proposed method for further use.
from the drawback of overfitting and also introduce the bias.
This method focuses on spatial property of signal not the
VI. C ONCLUSION
spectral one however, the proposed approach uses the spectral
property of feature, specially the beta and gamma frequency This article studies motor imagery task classification with
based features that are more related to the MI events. The beta and gamma frequency based sub-bands. These features
literature also suggests that event related synchronization are found to be very efficient for classification of motor
and desynchronization happens in the such situations. The imagery task. Initially 0.5 - 100Hz band pass filter is used
statistics of those feature provide more clear information to make the signal compatible with frequency bands of brain
than only wavelet based feature or PCA as compared in signal with 5 level DWT of dB4 wavelet. The Wavelets coef-
table VIII. The HS optimization technique helps to get the ficients of choice of frequency bands are extracted and further
better combination of feature set out of those features which statistics of these are calculated. These extracted features
further improve the performance and reduce the feature space. outperforms than other features. Further feature selection has
There are a number of techniques for classification of motor employed to find the relevant features. This method improves
imagery tasks but the proposed algorithm works with less the classification accuracy using weighted KNN. This shows
number of features and it also opens the another dimen- that HS based optimal feature selection method improves the
sion of research in MI task classification using optimization overall performance of motor imagery task classification. The
technique. DWT based features under the influence of gamma frequency
Apart from comparision, statistical analysis has been have measurable impact in motor imagery classification task.
performed to analyze the significance of proposed The proposed method in this article for feature extraction and
algorithm. The statistical analysis presented in further selection can be improved further with parameter tuning for
subsection. better classification accuracy of MI task. Gamma frequency

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 14270 IEEE SENSORS JOURNAL, VOL. 21, NO. 13, JULY 1, 2021

for MI task will also be further explore to find the dominance [21] H. Peng, F. Long, and C. Ding, “Feature selection based on mutual
of specific frequency band. information criteria of max-dependency, max-relevance, and min-
redundancy,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 8,
pp. 1226–1238, Aug. 2005.
R EFERENCES [22] G. Chandrashekar and F. Sahin, “A survey on feature selection methods,”
Comput. Electr. Eng., vol. 40, no. 1, pp. 16–28, Jan. 2014.
[1] J. R. Wolpaw, “Brain-computer interfaces as new brain output path-
[23] K. Kira and L. A. Rendell, “A practical approach to feature selection,” in
ways,” J. Physiol., vol. 579, no. 3, pp. 613–619, Mar. 2007.
Machine Learning Proceedings. Amsterdam, The Netherlands: Elsevier,
[2] I. Soumya, M. Zia Ur Rahman, D. V. R. K. Reddy, and A. Lay-Ekuakille,
1992, pp. 249–256.
“Efficient block processing of long duration biotelemetric brain data for
health care monitoring,” Rev. Sci. Instrum., vol. 86, no. 3, Mar. 2015, [24] X. Jin, A. Xu, R. Bie, and P. Guo, “Machine learning techniques
Art. no. 035003. and chi-square feature selection for cancer classification using SAGE
[3] G. V. S. Karthik, S. Y. Fathima, M. Z. U. Rahman, S. R. Ahamed, and gene expression profiles,” in Data Mining for Biomedical Applica-
A. Lay-Ekuakille, “Efficient signal conditioning techniques for brain tions. BioDM (Lecture Notes in Computer Science), vol. 3916, J.
activity in remote health monitoring network,” IEEE Sensors J., vol. 13, Li, Q. Yang, A. H. Tan, Eds. Berlin, Germany: Springer, 2006, doi:
no. 9, pp. 3276–3283, Sep. 2013. 10.1007/11691730_11.
[4] J. Wang, Z. Feng, N. Lu, L. Sun, and J. Luo, “An information fusion [25] J. Zhang, M. Chen, S. Zhao, S. Hu, Z. Shi, and Y. Cao, “ReliefF-
scheme based common spatial pattern method for classification of motor based EEG sensor selection methods for emotion recognition,” Sensors,
imagery tasks,” Biomed. Signal Process. Control, vol. 46, pp. 10–17, vol. 16, no. 10, p. 1558, Sep. 2016.
Sep. 2018. [26] A. R. Subhani, W. Mumtaz, N. Kamil, N. M. Saad, N. Nandagopal,
[5] C. A. Stefano Filho, R. Attux, and G. Castellano, “Can graph metrics and A. S. Malik, “MRMR based feature selection for the classification
be used for EEG-BCIs based on hand motor imagery?” Biomed. Signal of stress using EEG,” in Proc. 11th Int. Conf. Sens. Technol. (ICST),
Process. Control, vol. 40, pp. 359–365, Feb. 2018. Dec. 2017, pp. 1–4.
[6] B. Blankertz, R. Tomioka, S. Lemm, M. Kawanabe, and K.-R. Müller, [27] J. González, J. Ortega, M. Damas, P. Martín-Smith, and J. Q. Gan,
“Optimizing spatial filters for robust EEG single-trial analysis,” IEEE “A new multi-objective wrapper method for feature selection—Accuracy
Signal Process. Mag., vol. 25, no. 1, pp. 41–56, 2008. and stability analysis for BCI,” Neurocomputing, vol. 333, pp. 407–418,
[7] T.-E. Kam and S.-W. Lee, “Time-dependent common spatial patterns Mar. 2019.
optimization for EEG signal classification,” in Proc. 1st Asian Conf. [28] J. Ortega, J. Asensio-Cubero, J. Q. Gan, and A. Ortiz, “Classification
Pattern Recognit., Nov. 2011, pp. 643–646. of motor imagery tasks for BCI with multiresolution analysis and
[8] K. Keng Ang, Z. Yang Chin, H. Zhang, and C. Guan, “Filter bank multiobjective feature selection,” Biomed. Eng. OnLine, vol. 15, no. S1,
common spatial pattern (FBCSP) in brain-computer interface,” in Proc. p. 73, Jul. 2016.
IEEE Int. Joint Conf. Neural Netw. (IEEE World Congr. Comput. Intell.), [29] S.-L. Wu et al., “Fuzzy integral with particle swarm optimization for
Jun. 2008, pp. 2390–2397. a motor-imagery-based brain–computer interface,” IEEE Trans. Fuzzy
[9] B. Blankertz, M. Kawanabe, R. Tomioka, F. Hohlefeld, K.-R. Müller, Syst., vol. 25, no. 1, pp. 21–28, Feb. 2017.
and V. V. Nikulin, “Invariant common spatial patterns: Alleviating [30] M. Z. Baig, N. Aslam, H. P. H. Shum, and L. Zhang, “Differential
nonstationarities in brain-computer interfacing,” in Proc. Adv. Neural evolution algorithm as a tool for optimal feature subset selection in motor
Inf. Process. Syst., 2008, pp. 113–120. imagery EEG,” Expert Syst. Appl., vol. 90, pp. 184–195, Dec. 2017.
[10] F. Lotte and C. Guan, “Regularizing common spatial patterns to improve [31] Z. Woo Geem, J. Hoon Kim, and G. V. Loganathan, “A new heuristic
BCI designs: Unified theory and new algorithms,” IEEE Trans. Biomed. optimization algorithm: Harmony search,” Simulation, vol. 76, no. 2,
Eng., vol. 58, no. 2, pp. 355–362, Feb. 2011. pp. 60–68, Feb. 2001.
[11] H. Wang and X. Li, “Regularized filters for L1-norm-based common [32] P. Herman, G. Prasad, T. M. McGinnity, and D. Coyle, “Comparative
spatial patterns,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 24, no. 2, analysis of spectral approaches to feature extraction for EEG-based
pp. 201–211, Feb. 2016. motor imagery classification,” IEEE Trans. Neural Syst. Rehabil. Eng.,
[12] G. Pfurtscheller and A. Aranibar, “Event-related cortical desynchro- vol. 16, no. 4, pp. 317–326, Aug. 2008.
nization detected by power measurements of scalp EEG,” Elec- [33] D. Steyrl, R. Scherer, O. Förstner, and G. R. Müller-Putz, “Motor
troencephalogr. Clin. Neurophysiol., vol. 42, no. 6, pp. 817–826, imagery brain-computer interfaces: Random forests vs regularized lda-
Jun. 1977. non-linear beats linear,” in Proc. 6th Int. Brain-Comput. Interface Conf.,
[13] C. Park, D. Looney, N. ur Rehman, A. Ahrabian, and D. P. Mandic, Sep. 2014, pp. 241–244.
“Classification of motor imagery BCI using multivariate empirical mode [34] D. Steyrl, R. Scherer, J. Faller, and G. R. Müller-Putz, “Random forests
decomposition,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 21, no. 1, in non-invasive sensorimotor rhythm brain-computer interfaces: A practi-
pp. 10–22, Jan. 2013. cal and convenient non-linear classifier,” Biomed. Eng. / Biomedizinische
[14] J. Kevric and A. Subasi, “Comparison of signal decomposition Technik, vol. 61, no. 1, pp. 77–86, Feb. 2016.
methods in classification of EEG signals for motor-imagery BCI [35] S. Burrus, C. S. Burrus, R. A. Gopinath, and H. Guo, Introduction to
system,” Biomed. Signal Process. Control, vol. 31, pp. 398–406, Wavelets and Wavelet Transforms: A Primer. Upper Saddle River, NJ,
Jan. 2017. USA: Prentice-Hall, 1998.
[15] H. Mirvaziri and Z. S. Mobarakeh, “Improvement of EEG-based
[36] C. E. Heil and D. F. Walnut, “Continuous and discrete wavelet trans-
motor imagery classification using ring topology-based particle swarm
forms,” SIAM Rev., vol. 31, no. 4, pp. 628–666, 1989.
optimization,” Biomed. Signal Process. Control, vol. 32, pp. 69–75,
[37] A. Lay-Ekuakille et al., “Entropy index in quantitative EEG measure-
Feb. 2017.
ment for diagnosis accuracy,” IEEE Trans. Instrum. Meas., vol. 63, no. 6,
[16] J. Luo, Z. Feng, and N. Lu, “Spatio-temporal discrepancy feature for
pp. 1440–1450, Jun. 2014.
classification of motor imageries,” Biomed. Signal Process. Control,
vol. 47, pp. 137–144, Jan. 2019. [38] R. Chaurasiya, N. Londhe, and S. Ghosh, “Statistical wavelet features,
[17] K. Samanta, S. Chatterjee, and R. Bose, “Cross-subject motor imagery pca, and svm based approach for eeg signals classification,” Int. J.
tasks EEG signal classification employing multiplex weighted visibility Electr., Comput., Electron. Commun. Eng., vol. 9, no. 2, pp. 182–186,
graph and deep feature extraction,” IEEE Sensors Lett., vol. 4, no. 1, 2015.
pp. 1–4, Jan. 2020. [39] I. Guyon and A. Elisseeff, “An introduction to variable and feature
[18] L. Cheng, D. Li, G. Yu, Z. Zhang, X. Li, and S. Yu, “A motor selection,” J. Mach. Learn. Res., vol. 3, pp. 1157–1182, Jan. 2003.
imagery EEG feature extraction method based on energy principal [40] R. Diao and Q. Shen, “Feature selection with harmony search,” IEEE
component analysis and deep belief networks,” IEEE Access, vol. 8, Trans. Syst., Man, Cybern. B, Cybern., vol. 42, no. 6, pp. 1509–1523,
pp. 21453–21472, 2020. Dec. 2012.
[19] Y. Park and W. Chung, “Optimal channel selection using correlation [41] P.-N. Tan, Introduction to Data Mining. London, U.K.: Pearson, 2018.
coefficient for CSP based EEG classification,” IEEE Access, vol. 8, [42] N. Akanksha, S. Mridu, V. Shrish, and V. Raj, “Dimensionality reduction
pp. 111514–111521, 2020. for motor imagery signal classification using wavelet analysis,” Int. J.
[20] J. C. Ang, A. Mirzal, H. Haron, and H. N. A. Hamed, “Supervised, Control Theory Appl, vol. 10, pp. 65–76, 2017.
unsupervised, and semi-supervised feature selection: A review on gene [43] M. Sahu, Y. Sharma, D. Sharma, and S. Bajpai, “Feature compression
selection,” IEEE/ACM Trans. Comput. Biol. Bioinf., vol. 13, no. 5, using PCA on motor imagery classifications,” in Proc. 3rd Int. Conf.
pp. 971–989, Sep. 2016. Internet Things Connected Technol. (ICIoTCT), Apr. 2018, pp. 26–27.

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.
 MOHDIWALE et al.: STATISTICAL WAVELETS WITH HARMONY SEARCH-BASED OPTIMAL FEATURE SELECTION 14271

[44] M. Sahu and S. Shukla, “Impact of feature selection on EEG based G. R. Sinha (Senior Member, IEEE) is cur-
motor imagery,” in Information and Communication Technology for rently an Adjunct Professor with the Inter-
Competitive Strategies (Lecture Notes in Networks and Systems), vol. national Institute of Information Technology
40, S. Fong, S. Akashe, and P. Mahalle, Eds. Singapore: Springer, 2019, Bangalore (IIITB) and deputed as a Profes-
doi: 10.1007/978-981-13-0586-3_73. sor with the Myanmar Institute of Informa-
[45] J. Demšar, “Statistical comparisons of classifiers over multiple data sets,” tion Technology (MIIT), Mandalay, Myanmar.
J. Mach. Learn. Res., vol. 7, pp. 1–30, Jan. 2006. He is a Visiting Professor (Honorary) with
[46] P. Nemenyi, “Distribution-free multiple comparisons,” in Biometrics, Sri Lanka Technological Campus Colombo for
vol. 18, no. 2. Washington, DC, USA: International Biometric Society, one year from 2019 to 2020. He has more
1962, p. 263. than 200 research articles, edited books, and
books into his credit. He has edited books for
reputed International publishers. He is an active reviewer and an editorial
member of more than 12 reputed International Journals in his research
areas, such as the IEEE TRANSACTIONS, Elsevier Journals, and Springer
Journals. He has teaching and research experience of 21 years. He has
been the Dean of Faculty and Executive Council Member of CSVTU and
a member of Senate of MIIT. Dr. Sinha has been delivering ACM lectures
as an ACM Distinguished Speaker in the field of DSP since 2017 across
Samrudhi Mohdiwale received the bachelor’s the world. He was a recipient of many awards and recognitions at national
(Hons.) degree in engineering in electronics and and international level. He has delivered more than 50 Keynote/Invited
telecommunication engineering from the Shri Talks and Chaired many Technical Sessions in International Conferences
Shankaracharya Institute of Professional Man- across the world. He has supervised eight Ph.D. Scholars, 15 M.Tech.
agement and Technology at Raipur, in 2015, and Scholars, and has been Supervising one more Ph.D. Scholar. His
the Masters of Technology degree in information research interests include biometrics, cognitive science, medical image
technology from the National Institute of Tech- processing, computer vision, outcome-based education (OBE), and ICT
nology (NIT) at Raipur, Raipur, in 2018, where tools for developing employability skills. He is a Fellow of the Institute of
she is currently pursuing the Ph.D. degree. She Engineers India and Fellow of IETE India. He served as a Distinguished
has three years of research experience and in IEEE Lecturer in IEEE India council for Bombay section.
such a short span of time. She has published
more than seven research articles in the areas of brain computing, signal
processing, and machine learning. Her research interests include neuro-
computing, brain–computer interface, biomedical image, and signal Vikrant Bhateja (Senior Member, IEEE) is an
processing with machine learning and optimization techniques. Associate Professor with the Department of Elec-
tronics and Communication Engineering (ECE),
Shri Ramswaroop Memorial Group of Profes-
sional Colleges (SRMGPC), Lucknow, and also
the Dean (Academics and IQAC) with SRMGPC.
He is a doctorate in ECE (bio-medical imag-
ing) with a total academic teaching experi-
ence of 17 years with around 160 publications
in reputed international conferences, journals,
and online book chapter contributions, out of
Mridu Sahu received the degree in computer which 20 articles are published in SCIE indexed journals, with Scopus
science and engineering from the Maulana Azad h-index of 24. His research interests include digital image and video
National Institute of Technology, Bhopal, in 2004, processing, computer vision, medical imaging, and machine learning.
the Master of Technology degree in computer sci- He has been instrumental in chairing/co-chairing around 25 interna-
ence and engineering from RIT Raipur, Raipur, tional conferences in India and abroad as Publication/TPC chair and
India, in 2011, and the Ph.D. degree in computer edited 26 book volumes from Springer-Nature as a corresponding/co-
science and engineering from the National Insti- editor/author on date. He has been the Editor-in-Chief of the IGI Global–
tute of Technology at Raipur, Raipur, in 2018. She International Journal of Natural Computing and Research (IJNCR) since
has more than ten year experiences in teaching. 2017. He has guest edited many special issues in reputed Scopus/SCIE
She is working as an Assistant Professor with indexed journals including Evolutionary Intelligence and Arabian Journal
the Department of Information Technology, NIT of Science and Engineering under Springer-Nature. He has received the
Raipur. She has published more than 25 research articles in various Certificate of Outstanding Contribution in Reviewing from 2016 to 2017 in
journals and conferences and book chapters in the field of data mining, prestigious Elsevier journals like: AEUE, Measurements, and Computer
brain computer interface, sensor devices, and visual mining techniques. Methods and Programs in Bio-medicine.

Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on November 07,2024 at 08:23:30 UTC from IEEE Xplore. Restrictions apply.