Professional Documents
Culture Documents
Features based on analytic IMF for classifying motor imagery EEG signals in
BCI applications
PII: S0263-2241(17)30702-9
DOI: https://doi.org/10.1016/j.measurement.2017.10.067
Reference: MEASUR 5069
Please cite this article as: S. Taran, V. Bajaj, D. Sharma, S. Siuly, A. Sengur, Features based on analytic IMF for
classifying motor imagery EEG signals in BCI applications, Measurement (2017), doi: https://doi.org/10.1016/
j.measurement.2017.10.067
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Features based on analytic IMF for classifying motor
imagery EEG signals in BCI applications
Melbourne, Australia
c Firat University, Technology Faculty, Electrical and Electronics Engineering Dept., Elazig,
TURKEY
Abstract
∗ Corresponding author
Email address: sachin.taran@iiitdmj.ac.in, bajajvarun056@yahoo.co.in,
dheeraj@iiitdmj.ac.in, siuly.siuly@vu.edu.au, ksengur@gmail.com (Sachin Tarana ,
Varun Bajaja , Dheeraj Sharmaa , Siuly Siulyb , A. Sengurc )
1. Introduction
2
for classification. The multivariate decomposition and short time Fourier trans-
form (FT) analyzes the EEG signals and magnitude spectrum based features
classify the MI tasks EEG signals [14]. The Dual tree complex WT decomposed
30 EEG signals into different bands of real and imaginary coefficients. Energy of
respective bands coefficients are considered as features and tested on several
classification algorithms to classify left and right hand MI tasks EEG signals
[15].
The wavelet packet decomposition decomposed de-noising EEG signal into
35 sub-bands, which are used in the extraction of higher order statistical features
[16]. These features are fed as inputs to K-nearest neighbour classifier and
results 92.8% classification accuracy [16]. An optimum allocation based statis-
tical approach is applied for reduction in length of EEG sequences and these
sequences are used in the extraction of statistical features [17]. The extracted
40 features are applied as inputs to naive base classifier, which results in 96.3% ac-
curacy and 2.32% standard deviation (SD). The common spatial pattern (CSP)
method is used for feature extraction and z-score linear discriminant analy-
sis (Z-LDA) classifier provides the 81.1% classification accuracy and 18.2% SD
[18]. The cross-correlation based approach is proposed for extraction of statis-
45 tical features, which are fed as inputs to LS-SVM classifier and the outcome
of methodology came up in form of 95.72% accuracy and 4.35% SD [19]. The
regularized-CSP with aggregation is proposed for classification of MI based EEG
signals, which provides 83.9% classification accuracy and 10.36% SD [20]. As
observed from the literature survey, no one has been applied analytic intrinsic
50 mode functions (AIMFs) based feature extraction method for classification of
MI tasks and BCI system can be improved for identification of different MI
tasks.
In this work, AIMFs based features are proposed for automatic classification
of MI tasks dependent EEG signals. The multichannel recorded EEG signals
55 show non-linear and non-stationary nature. The complexity of such EEG sig-
nals is explored by AIMFs-based features. The AIMFs from raw EEG signal are
obtained by applying empirical mode decomposition (EMD) and Hilbert trans-
3
form (HT). The features based on AIMFs are: raw moment of the first derivative
of instantaneous frequency (RMFDIF), area, spectral moment of power spec-
60 tral density (SMPSD), and peak value of PSD. Further, the unit length feature
normalization method is applied on extracted features to performed unbiased
classification. The normalized features are applied as inputs to LS-SVM classi-
fier to evaluate the MI-tasks classification performance of proposed method.
The sequence of remaining paper is as follows. The proposed methodology
65 described in Section 2, under the following subsections: data set, decomposition
technique, AIMF based features extraction, feature normalization, and classi-
fication. Section 3 presents the results of proposed method and discussion is
given in Section 4. The conclusion of paper is given in Section 5.
2. Methodology
2.1. Dataset
The EMD decomposes non-stationary EEG signal into stationary basis func-
tions which are called as intrinsic mode functions (IMFs) [22]. These IMFs have
to satisfy two basic characteristics. (i) In the complete data set, the count of
maximum and minimum values must be equal or differ by at most one from the
85 number of baseline crossing. (ii) The mean value of envelopes defined by local
4
Figure 1: EEG based MI tasks classification process.
maxima and minima are zero. First condition improves, the stationary char-
acteristics and second improves, the definition of instantaneous frequency (IF).
After applying EMD, the non-stationary EEG signal x[n] can be decomposed
into M IMFs as:
M
X
x[n] = cm [n] + rM [n] (1)
m=1
90 where, cm [n] is the mth IMF and rM [n] is final residue. The RH MI task EEG
signal and its IMFs are shown in Figure 2.
Hilbert transform (HT) provides the analytical representation of IMF, which
is an extension of real-valued IMF to complex valued IMF. The analytic IMF
can be represented as [22]:
95 where, cH [n] is the HT of IMF c[n], A[n] is time dependent amplitude, and φ[n]
is instantaneous phase of a AIMF.
5
Figure 2: The IMFs of right hand (RH) MI task EEG signal.
The AIMFs assessed the following characteristics of EEG signals. (i) The
AIMFs characterizes the time-dependent amplitude and frequency component
100 corresponding to the variation of EEG signal. (ii) The AIMFs of EEG signal
rotate in a specific direction around the unique center. (iii) AIMFs have multi-
resolution property, so proposed features are able to extract different levels of
information from EEG signals. In this work following AIMFs based features are
extracted for classification of different MI-tasks EEG signals.
6
of mth AIMF (Φm ) and its difference (δf ) is defined as:
1
fi = dif f {Φm }, δf = |dif f (fi )|. (3)
2π
110 The IFs of IMFs of Figure 2 is shown in Figure 3. IF difference (δf ) defined in
Eq. (3), is used for the computation of RMFDIF feature as [23]:
N −1
1 X
RM F DIF = δf [n] (4)
N − 1 n=1
7
0.5
1 if [<{z[n]}]2 + [={z[n]}]2 <r 123
δd [n] = (6)
0 otherwise
where, N is a total number of samples in AIMF and the area of AIMF computed
by the radius (r) of circular region to define 95% CTM.
125 The spectral moment of PSD is used to define greater order shape of EEG data
points, which can be defined as [25]:
L
X
SM P SD = k.P SDk (8)
k=1
8
where, L is the number of point in PSD.
9
a11 ∗ w1q a12 ∗ w1q . . . . a1q ∗ w1q
a21 ∗ w2q a22 ∗ w2q a2q ∗ w2q
. . . .
FN =
. . . . . . .
ap1 ∗ wpq ap2 ∗ wpq . . . .apq ∗ wpq
where, FN is unit length normalized feature matrix.
150 The proposed AIMFs based features are tested on LS-SVM classifier for the
classification of RH and RF MI tasks EEG signals. The final decision function
for the LSSVM classifier is given as [28]:
"N #
X T
3. Results
10
of Kruskal-Wallis statistical test. Tables 1-4 present, the p−values of extracted
features RMFDIF, area, SMPSD, and peak value of PSD, respectively. In these
170 tables, the lesser p−values of all features show their statistical significance for
discrimination of different MI tasks. In Tables 1-4, higher frequency IMFs based
features are more statistically significant for all channels as compared to lower
frequency IMFs based features. Hence, first four IMFs-based features are consid-
ered for the classification of different MI tasks. The ranges (mean and standard
175 deviation) of these IMFs-based features for RH and RF MI tasks EEG signals
are presented in Tables 5-8. Table 5 shows that, the range of RMFDIF feature
is higher for RH MI task EEG signals as compared to RF MI task EEG signals,
for all IMFs. This range variation might be shown because of the larger rate of
change of IF corresponding to RH MI task EEG signals as compared to RF MI
180 tasks EEG signals. Table 6 presents the variation of area feature, the IMF1 and
IMF2 based area feature values show larger variation corresponding to RF MI
task EEG signals. It could be due to, the larger amplitude of IMF1 and IMF2
of RF MI task EEG signals as compared to RH MI task EEG signals. The
opposite behaviour is observed for IMF3 and IMF4, in Table 6. The variation
185 in SMPSD feature is presented in Table 7, the SMPSD shows higher values for
IMF1 and IMF2 of RF MI task EEG signals. It might be happen due to the
larger amplitude of IMF1 and IMF2 of RF MI task EEG signals as compared to
RH MI task EEG signals. The opposite nature is observed in Table 7, for IMF3
and IMF4. The peak value of PSD is simply the maximum overshoot of PSD,
190 which depends on the maximum value of IMF. In Table 8, the larger values of
the peak of PSD feature observed corresponding to RH MI task EEG signals.
It could be due to the larger maximum value of IMFs corresponding to RH MI
task EEG signals as compared to RF MI task EEG signals.
The proposed features in Tables 1-8 clearly show the differences in electrical
195 activity of brain corresponding to RH and RF MI tasks, EEG signals. The
dimension of IMFs-wise feature-sets, for RH and RF MI tasks EEG signals are
40 × 200 and 40 × 250, respectively. In these dimensions, 40 rows represent the
four features extracted for 10-channels, while 200 and 250 columns represent
11
the number of observation of RH and RF MI tasks EEG signals, respectively.
200 The IMF-wise feature set is applied as input to LS-SVM classifier for classifi-
cation of RH and RF MI tasks, EEG signals. The classification performance
of LS-SVM classifier is evaluated in terms of following performance parameters:
accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value
(PPV), negative predictive value (NPV), and error rate detection (ERD). These
205 parameters are computed using 10-cross validation approach and different ker-
nel functions for LS-SVM classifier. In this work, the used kernel functions and
their parameter are as follows: linear kernel is parameter independent, polyno-
mial kernel of degree l=5, RBF kernel of width σ=10. Table 9 presents, the
IMFs wise summary of evaluated performance parameters with different Kernel
210 functions of the LS-SVM classifier. The overall best performance parameters are
obtained for IMF1 with RBF kernel in Table 9 as: ACC 97.56%, SEN 96.45%,
SPE 98.96%, PPV 99.2%, and NPV 95.2%, and ERD 4.28%.
4. Discussions
12
Table 1: The p−values of RMFDIF feature for different IMFs.
Channel1 9.75 ∗ 10−06 1.54 ∗ 10−12 8.89 ∗ 10−07 7.18 ∗ 10−05 0.0024 0.0588
Channel3 0.0286 2.44 ∗ 10−13 1.85 ∗ 10−09 7.05 ∗ 10−07 0.2929 0.0875
Channel4 0.0016 5.48 ∗ 10−17 1.3 ∗ 10−08 9.89 ∗ 10−05 0.0758 0.4047
Channel6 6.89 ∗ 10−19 2.36 ∗ 10−14 1.14 ∗ 10−07 0.0002 0.053 0.0163
Channel7 0.0008 7.43 ∗ 10−18 1.08 ∗ 10−10 1.8 ∗ 10−05 0.0913 0.0105
Channel8 4.35 ∗ 10−05 2.94 ∗ 10−19 3.25 ∗ 10−12 2.59 ∗ 10−06 0.0018 0.0095
Channel9 4.61 ∗ 10−11 2.06 ∗ 10−19 1.49 ∗ 10−07 8.16 ∗ 10−06 0.0013 0.0256
Channel10 3.21 ∗ 10−10 2.52 ∗ 10−18 3.31 ∗ 10−12 3.92 ∗ 10−57 0.0472 0.001
Channel1 1.4 ∗ 10−48 1.6 ∗ 10− 19 1.97 ∗ 10−77 0.0014 4.34 ∗ 10−40 0.001
Channel2 2.36 ∗ 10−10 0.4199 2.77 ∗ 10−35 8.24 ∗ 10−13 3.53 ∗ 10−41 0.7949
Channel3 1.14 ∗ 10−08 0.002 7.51 ∗ 10−63 9.89 ∗ 10−09 0.6643 0.0032
Channel4 1.44 ∗ 10−12 0.0005 1.21 ∗ 10−06 5.83 ∗ 10−19 0.9388 0.0003
Channel5 1.95 ∗ 10−10 8.31 ∗ 10−05 1.66 ∗ 10−25 0.0088 0.3345 0.7772
Channel7 6.38 ∗ 10−12 2.91 ∗ 10−08 2.07 ∗ 10−29 5.39 ∗ 10−17 1.13 ∗ 0−06 9.18 ∗ 10−64
Channel8 4.41 ∗ 10−82 4.41 ∗ 10−82 4.41 ∗ 10−82 1.75 ∗ 10−81 2.2 ∗ 10−80 8.37 ∗ 10−71
Channel9 0.1549 7.08 ∗ 10−19 2.19 ∗ 10−11 9.57 ∗ 10−23 1.75 ∗ 10−12 0.0006
Channel10 1.92 ∗ 10−14 3.71 ∗ 10−09 3.54 ∗ 10−57 2.96 ∗ 10−07 4.49 ∗ 10−44 0.0008
13
Table 3: The p−values of SMPSD feature for different IMFs.
Channel1 6.53 ∗ 10−82 4.64 ∗ 10−65 8.26 ∗ 10−78 1.03 ∗ 10−59 1.36 ∗ 10−07 0.6416
Channel2 0.4061 0.0354 8.39 ∗ 10−46 5.59 ∗ 10−34 0.1894 1.14 ∗ 10−08
Channel3 2.17 ∗ 10− 21 0.089 4.73 ∗ 10−08 5.33 ∗ 10−67 0.6933 3.46 ∗ 10−27
Channel4 4.69 ∗ 10−07 2.27 ∗ 10−05 0.0041 9.61 ∗ 10−54 2.47 ∗ 10−09 7.1 ∗ 10−08
Channel5 2.04 ∗ 10−60 2.32 ∗ 10−09 0.0004 7.85 ∗ 10−80 0.1713 5.84 ∗ 10−08
Channel7 0.0002 1.07 ∗ 10−27 4.1 ∗ 10−17 2.36 ∗ 10−50 1.88 ∗ 10−10 1.57 ∗ 10−10
Channel8 4.41 ∗ 10−82 4.41 ∗ 10−82 1.12 ∗ 10−80 2.4 ∗ 10−62 4.41 ∗ 10−82 6.46 ∗ 10−82
Channel9 1.26 ∗ 10−25 8.32 ∗ 10−19 2.82 ∗ 10−61 2.16 ∗ 10−77 0.0022 9.55 ∗ 10−08
Channel10 2.61 ∗ 10−58 3.53 ∗ 10−09 6.43 ∗ 10−63 0.0007 0.0487 0.0472
Channel1 3.17 ∗ 10−41 4.18 ∗ 10−70 0.007 8.84 ∗ 0−07 5.29 ∗ 10−23 3.11 ∗ 10−55
Channel2 1.93 ∗ 10−51 3.53 ∗ 10−36 1.24 ∗ 10−06 1.8 ∗ 10−09 0.0009 0.3332
Channel4 0.038 5.21 ∗ 10−37 2.84 ∗ 10−13 2.44 ∗ 10−17 0.0008 1.04 ∗ 10−6
Channel7 8.72 − 10−20 2.24 ∗ 10−31 4.9 ∗ 10−05 2.44 ∗ 10−08 0.0003 0.0038
Channel8 8.57 ∗ 10−81 1.59 ∗ 10−69 4.4 ∗ 10−82 4.52 ∗ 10−82 4.4 ∗ 10−82 4.4 ∗ 10−82
Channel9 8.53 ∗ 10−34 2.1 ∗ 10−78 2.08 ∗ 10−15 9.54 ∗ 10−26 0.0013 0.004
14
Table 5: Ranges (Mean±Standard Deviation) of RMFDIF feature for IMFs of RH and RF
MI task EEG signals.
15
Table 6: Ranges of area feature for IMFs of RH and RF MI task EEG signals.
16
Table 7: Ranges of SMPSD feature for IMFs of RH and RF MI task EEG signals.
17
Table 8: Ranges of peak of PSD feature for IMFs of RH and RF MI task EEG signals.
18
Table 9: The performance parameters of proposed method using different kernels of LS-SVM
classifier.
19
Table 10: The performance report of proposed method with other existing methods in terms
of subjects (aa, al, av, aw, ay) wise accuracy (%), average accuracy (AACC), and standard
deviation (SD).
Proposed Approach AFE and FN based approach 97.78 98.89 96.67 98.89 95.56 97.56 1.45
Kevric et al. [16] MSPCA, WPD, HOS and k-NN 96 92.3 88.9 95.4 91.4 92.8 2.93
Wang et al. [17] OA and NB based approach 97.92 97.88 98.26 94.47 93.26 96.36 2.32
Zhang et al. [18] Z-LDA 77.7 100.0 68.4 99.60 59.9 81.1 18.2
Siuly and Li [19] CC based LS-SVM 97.88 99.17 98.75 93.43 89.36 95.72 4.35
Lu et al. [20] R-CSP with aggregation 76.80 98.20 74.50 92.90 77.00 83.90 10.86
Where, AFE analytic feature extraction and FN feature normalization, MSPCA multi-scale principal component
analysis, WPD wavelet packet decomposition, HOS higher order statistics, and K-NN K-nearest neighbor, OA op-
timum allocation and NB naive bayes, Z-LDA Z-score linear discriminant analysis, CC cross correlation, R-CSP
regularized common spatial pattern.
nals. Figure 5 clearly shows that both class detection rates are closer to highest
100% detection rate. Hence, Table 9 and Figure 5 evident that proposed method
230 provides a fairly accurate and stable solution, for the classification of RH and
RF MI tasks, EEG signals. Table 10 presents, the performance comparison re-
port of proposed method with other existing methods. The proposed method
provides consistent subjects-wise performance in a range of 95.56-98.89%, while
the other methods show fluctuating results in subject-wise performance, which
235 is also reflected in terms of their larger value of standard deviation. In Table 10,
the proposed method achieved highest MI tasks classification accuracy 97.56%
and minimum standard deviation 1.45% as compared to other existing methods.
The results of proposed method show its deftness for classification of different
MI tasks in BCI system. Therefore, the proposed method can be used for devel-
240 oping practical application of RH and RF MI tasks in BCI systems, like robotic
arm and wheelchair controlling etc.
20
5. Conclusion
260 References
21
nals using tunable-Q factor wavelet transform and bootstrap aggregating,
270 Computer methods and programs in biomedicine 137 (2016) 247–259.
22
295 [11] H.-I. Suk, S.-W. Lee, A novel bayesian framework for discriminative fea-
ture extraction in brain-computer interfaces, IEEE Transactions on Pattern
Analysis and Machine Intelligence 35 (2) (2013) 286–299.
[12] M. Li, C. Lu, The recognition of EEG with CSSD and SVM, in: Intelligent
Control and Automation (WCICA), 2012 10th World Congress on, IEEE,
300 2012, pp. 4741–4746.
[17] H. Wang, Y. Zhang, et al., Detection of motor imagery EEG signals employ-
315 ing naı̈ve bayes based learning process, Measurement 86 (2016) 148–158.
[18] R. Zhang, P. Xu, L. Guo, Y. Zhang, P. Li, D. Yao, Z-score linear discrim-
inant analysis for EEG based brain-computer interfaces, PloS one 8 (9)
(2013) e74433.
[19] S. Siuly, Y. Li, Improving the separability of motor imagery EEG sig-
320 nals using a cross correlation-based least square support vector machine
for brain–computer interface, IEEE Transactions on Neural Systems and
Rehabilitation Engineering 20 (4) (2012) 526–538.
23
[20] H. Lu, H.-L. Eng, C. Guan, K. N. Plataniotis, A. N. Venetsanopoulos, Reg-
ularized common spatial pattern with aggregation for EEG classification in
325 small-sample setting, IEEE transactions on Biomedical Engineering 57 (12)
(2010) 2936–2946.
340 [24] R. B. Pachori, V. Bajaj, Analysis of normal and epileptic seizure EEG
signals using empirical mode decomposition, Computer methods and pro-
grams in biomedicine 104 (3) (2011) 373–381.
24
[28] J. A. Suykens, J. Vandewalle, Least squares support vector machine clas-
sifiers, Neural processing letters 9 (3) (1999) 293–300.
25
Highlights:
Features are extracted from analytic intrinsic mode functions of EEG signals.
Features used as input to LS-SVM classifier for classification of MI tasks EEG signals.
Proposed features with LS-SVM classifier provide better performance as compared to existing methods.