Professional Documents
Culture Documents
https://doi.org/10.1007/s00542-018-4075-z (0123456789().,-volV)(0123456789().
,- volV)
TECHNICAL PAPER
Abstract
EEG signals are non-stationary, complex and non-linear signals. During major depressive disorder (MDD) or depression,
any deterioration in the brain function is reflected in the EEG signals. In this paper, linear features (band power, inter
hemispheric asymmetry) and non-linear features [relative wavelet energy (RWE) and wavelet entropy (WE)] and com-
bination of linear and non-linear features were used to classify depression patients and healthy individuals. In this analysis
the data set used is publicly available data set contributed by Mumtaz et al. (Biomed Signal Process Control 31:108–115,
2017b). The dataset consisted of 34 MDD patients and 30 healthy individuals. The classifiers used were multi layered
perceptron neural network (MLPNN), radial basis function network (RBFN), linear discriminant analysis (LDA) and
quadratic discriminant analysis. When linear feature was used, highest classification accuracy of 91.67% was obtained by
alpha power with MLPNN classifier. When non-linear feature was used, both RWE and WE provided highest classification
accuracy of 90% with RBFN and LDA classifier, respectively. The highest classification of 93.33% was achieved when
combining linear and non-linear feature, i.e., combination alpha power and RWE with MLPNN as well as RBFN classifier.
This paper also showed that the combination of non-linear features, i.e., RWE and WE also performed the best with highest
classification accuracy of 93.33%. The study compared the accuracy, sensitivity and specificity of different classifiers along
with linear and non-linear features and combination of both. The results indicated that combination alpha power and RWE
showed the highest classification 93.33% accuracy in all the applied classifiers.
123
Microsystem Technologies
MDD causes degradation in brain’s performance which (Mohammadi et al. 2015). Wavelet transform features
is expected to be reflected in the bioelectrical activity of the when used in logistic regression (LR) provided a classifi-
brain. Screening of brain under these conditions with the cation accuracy of 87.5% (Mumtaz et al. 2017a, b). LDA
help of EEG can be helpful in understanding the brain along with detrended fluctuation analysis (DFA) and
functioning and imbalance of brain activity. EEG is cost Spectral Asymmetric Index (SASI) provided an accuracy
effective and also easily available. The time scale of EEG of 91.2% (Bachmann et al. 2017). LR, LDA and k-nearest
is in milli-seconds which is also same as that of neural neighbor (KNN) along with four non-linear features, i.e.,
activity thus giving better temporal resolution. DFA, Higuchi’s fractal dimension (HFD), correlation
Manual interpretation of the EEG signals is very com- dimension (CD) and Lyapunov exponent (LE) provided an
plicated as the EEG signals are extremely intricate, non- accuracy of 90, 86.6 and 80%, respectively. Both LR and
stationary and non-linear. So computer aided signal pro- LDA provided an accuracy of 73.3% while KNN provided
cessing becomes necessary to automatically classify MDD an accuracy of 70% when alpha power was used as a
patients and healthy individuals. feature (Hosseinifarda et al. 2013). Both SASI and HFD
Extensive research has been done using different EEG provided a classification accuracy of 85% using statistical
features and classifiers to improve the prediction accuracy t test (Bachmann et al. 2013).
of MDD patients. The purpose of this paper is to compare the efficiency of
It was found that during neuro-feedback treatment, linear features (band power and asymmetry), nonlinear
alpha asymmetry decreases remarkably in MDD patients features (WE and RWE) and combination of linear and
especially in women (Bruder et al. 2012). Frontal alpha non-linear features using different classifier (MLPNN,
asymmetry was found to be higher in MDD patients and RBFN, LDA, QDA) for detection of MDD.
found to be positively correlated with inhibitory behavior
before treatment (Gollan et al. 2014). Higher classification
accuracy in classifying depression and healthy individuals 2 Materials and methods
was achieved by alpha band power as compared to all the
bands that were analyzed together (Hosseinifarda et al. At first features are extracted from the EEG signals of 30
2013; Mohammadi et al. 2015). During depression, high MDD patients and 30 healthy individuals. The study is
alpha activity was found in the posterior region of the brain basically divided into three parts based on type of feature
in resting state (Stewart et al. 2014; Grin-Yatsenko et al. sets used. Three types of feature sets have been used: eight
2010). EEG power was found to be increased in central, linear feature sets (Four band power and four correspond-
occipital, parietal and posterior temporal areas in patients ing hemispheric asymmetry), two non-linear feature sets
in the early stage of depression (Grin-Yatsenko et al. (RWE and WE) and six combinations of linear and non-
2010). It was found that any of the hemispheres could be linear feature sets.
affected in depression but mainly abnormal sources were The linear feature sets for band power consisted of
found in the right hemisphere (Ricardo-Garcel et al. 2010). 19 9 1 equal to 19 features.
WE and approximate entropy was found to be higher in The non-linear feature sets consisted of 19 9 4 equal to
healthy individuals than MDD patients (Puthankattil and 76 features. The high dimensionality of non-linear feature
Joseph 2014). sets is reduced by dimension reduction with the help of
RWE for 4–32 Hz was found to be higher for healthy principal component analysis (PCA).
individuals than MDD patients in left hemisphere while For combination feature sets, whenever the combination
RWE for 0–4 Hz was found to be higher for MDD patients is with non-linear features dimension reduction is done
in right hemisphere of the brain (Puthankattil and Joseph with the help of PCA. Subsequently for each feature sets
2012). It was found that MDD patient’s beta power four types of classifiers are applied, i.e., MLPNN, RBFN,
increased only for left hemisphere (Grin-Yatsenko et al. LDA and QDA. Here, Fig. 1 represents the block diagram
2010; Mumtaz et al. 2017a, b). It was also found that beta of the proposed model for detection of MDD patients.
power was higher in central, temporal and parietal region
of brain for MDD patients (Mumtaz et al. 2017a, b). In a 2.1 Subjects
study it was found that while listening to music frontal
theta asymmetry was found to be decreased in MDD The EEG signal data set which has been used is publically
patients while frontal theta asymmetry was found to be available data contributed by Mumtaz et al. (2017b). The
increased in healthy individuals (Dharmadhikari et al. data sets consist of EEG signal of 34 MDD patients and 30
2018). healthy individuals. The group of 34 MDD patients con-
Classification accuracy of 80% was obtained using EEG sists of 17 male and 17 female patients with average age of
band power as feature and decision tree as classifier 40.3 ± 12.9 years. Out of 34 MDD patients EEG signal,
123
Microsystem Technologies
The EEG data was recorded for 5 min eyes closed (EC)
condition in resting position. The data was recorded with
19 electrodes based on international standard 10/20 system.
The 19 electrodes covered the frontal (Fp1, Fp2, F3, F4,
F7, F8, Fz), temporal (T3, T4, T5, T6), parietal (P3, P4,
Pz), occipital (O1, O2) central (C3, C4, Cz) as shown in
Fig. 2. The data was recorded using linked ear (LE)
Fig. 2 International 10/20 system for electrode positioning
123
Microsystem Technologies
2.4 Feature extraction alpha interhemispheric asymmetry, for theta power is ter-
med as theta asymmetry, for beta power is termed as beta
The feature extracted from the EEG signal can mainly be asymmetry. In this analysis, interhemispheric asymmetry
divided into two categories: linear analysis and non-linear was calculated for each pair of channels, i.e.,
analysis. Fp2Fp1; F4F3; F8F7; C4C3;
Frequency analysis (e.g. Fourier transform) and para-
T4T3; P4P3; T6T5; O2O1
metric modeling (auto regressive models) falls under linear
analysis. Linear methods include band power, Interhemi-
spheric asymmetry, EEG measurements (amplitude, fre- 2.4.3 Wavelet transform: non-linear features
quency, power) and so on. Linear methods have been
applied in a number of studies relating to EEG signal and For non-linear feature extraction, wavelet transform (WT)
have provided good results but it fails to capture the of the signal is done. Wavelet transform at low frequency
underlying complex and chaotic behavior of brain signals. gives good frequency information and at high frequency
Non-linear methods have the ability to capture the gives good time information. Wavelet transform provides
chaotic behavior and abrupt changes in EEG signal due to time frequency localization (Mallat 1989; Bopardikar and
underlying physiological phenomenon occurring in the Rao 1998). Wavelets are simple oscillatory function of
brain (Rodreguez-Bermudez and Garcia-Laencina 2015). fixed duration which by translation and dilation can create
Non-linear methods include relative wavelet energy any signal. CWT of finite energy function x(t) w.r.t to
(RWE), wavelet entropy, approximate entropy (ApEn), mother wavelet ws;d ðtÞ is represented as
Higuchi’s fractal dimension (HFD), correlation dimension
Z1
(CD), detrended fluctuation analysis (DFA) and so on.
In this paper, linear as well as non-linear methods have Ycwt ðs; dÞ ¼ xðtÞws;d ðtÞdt ð2Þ
been used. Linear methods used are band power and 1
123
Microsystem Technologies
pass filter and scaling function is related to low pass filter. 2.4.5 Wavelet entropy (WE)
Information of the wavelet coefficient is organized in a
multi resolution scheme represented as a hierarchal Information contained in each level of decomposition can
scheme. be represented as wavelet entropy (Rosso et al. 2006).
To obtain the original signal inverse filtering opera- Shanon entropy forms the basis of wavelet entropy.
tion is done. Output of low pass filter is called approx- The equation for calculation of Wavelet Entropy (WE)
imate (A) and output of the high pass filter is called is:
details (D). X
N
In this paper, three level decomposition using Coiflet 5 Ent ¼ RWEk logðRWEk Þ ð8Þ
was used as shown in Fig. 3. j¼1
123
Microsystem Technologies
Feature Vector ¼ ðeig1 ; eig2 . . .eigd Þ may become linearly separable in higher dimensional
space.
To get the final reduced dataset the following equation is The hidden layers perform non-linear transformation by
used using Radial Basis Function. The output layer uses linear
FinalDataSet ¼ RowFeatureVector transformation.
RowMeanAdjustData ð10Þ Each node in hidden layer is mathematically represented
by a radial basis function. Radial basis function is defined
where, RowFeatureVector represents the Eigen Vector in
as
row in descending order of their eigen values.
RowMeanAdjustData represents the mean adjusted data /j ðxÞ ¼ / x xj ð12Þ
where each row hold separate dimension
where, j = 1, 2,…, N and xj defines the center of the radial
The reduced dimension k should be selected such that
basis function.
90–95% of total variance is represented.
Generally, Gaussian radial function is used for the
Pk
ki computation in the hidden layer nodes which is defined as
V ¼ Pi¼1
d
0:9or0:95 ð11Þ
j¼1 kj /j ðxÞ ¼ /ðx xj Þ
!
1 2 ð13Þ
¼ exp 2 x xj
2.6 Classifier model 2rj
123
Microsystem Technologies
ðl2 l1 Þ2 TP
J¼ ð14Þ Sensitivity ¼ ð17Þ
r2 þ r1 ðTP þ FNÞ
2.6.4 Quadratic discriminant analysis (QDA) 3.1 Results for classification based on linear
methods
The decision boundary of QDA is non-linear (James 2013;
Tharwat 2016). QDA is very similar to LDA except that in 3.1.1 Band power
LDA the covariance are assumed to be same for each
classes but in QDA each class has its own co variance Figure 4 shows that among all the band powers alpha
matrix. power shows the highest accuracy of 91.67% in MLPNN.
LDA classification can be done on the basis of Baye’s The classification accuracy of all band powers is always
Theorem. A test sample x is said to belong to class k for found to be lesser than the accuracy of alpha power in all
which discriminant functions dk ðxÞ is maximum. types of applied classifier.
1 X
1
dk ðxÞ ¼ ðx lk Þ2 ðx lk Þ 3.1.2 Interhemispheric asymmetry
2 k
ð16Þ
1 X Figure 4 shows that alpha asymmetry has the highest
log þ log pk classification accuracy among all types of asymmetry in all
2 k
types of applied classifier. Highest classification accuracy
P
where k represents covariance matrix of class k. of 73.33% was shown in QDA for alpha asymmetry. Theta
When the covariance matrix of each of the classes is asymmetry classification accuracy was also found to be at
equal then LDA performs better than QDA but when the par with alpha asymmetry. Highest classification accuracy
covariance matrix differs QDA performs better than LDA. of theta asymmetry was found to be 71.67% in MLPNN.
Tables 1, 2, 3, 4 shows the performance of classifiers
2.7 Validation MLPNN, RBFN, LDA, QDA respectively using linear
features (Band power & Interhemispheric Asymmetry)
The classifiers performance was evaluated using tenfold only.
cross validation for 100 iterations. In this method the data
set is divided into ten equal size segments. Each nine 3.2 Results for classification based on non-linear
segments are randomly selected for training and one seg- features
ment is selected for testing. Validation is done in order to
validate the model on different data set other than the Figure 5 shows that on an average both RWE and WE has
training data set which was used for parameter estimation good classification accuracy. RWE and WE both showed
of the model. the highest classification accuracy of 90% with RBFN and
Accuracy, sensitivity and specificity are calculated on LDA classifiers, respectively. Tables 5, 6, 7, 8 shows the
the basis of confusion matrix performance of classifiers MLPNN, RBFN, LDA, QDA
respectively using non-linear features (RWE & WE) only.
123
Microsystem Technologies
Fig. 4 Comparison of
classification accuracy of linear
features
Table 1 MLPNN classifier’s performance for linear features Table 3 LDA classifier’s performance for linear features
Features Accuracy (%) Sensitivity (%) Specificity (%) Features Accuracy (%) Sensitivity (%) Specificity (%)
Table 2 RBFN classifier’s performance for linear features Table 4 QDA classifier’s performance for linear features
Features Accuracy (%) Sensitivity (%) Specificity (%) Features Accuracy (%) Sensitivity (%) Specificity (%)
3.3 Results for classification based RBFN and LDA. Combination of two non-linear features
on combination of linear and non-linear (RWE and WE) also performed well in almost all applied
features classifiers with highest classification accuracy of 93.33% in
LDA. Whenever the alpha asymmetry feature was com-
Figure 6 shows that highest classification accuracy of bined with the combination alpha power and RWE, accu-
93.33% was achieved by the combination of alpha power racy degraded in all the classifier. Tables 9, 10, 11,
(linear method) and RWE (non-linear method) in MLPNN, 12 shows the performance of classifiers MLPNN, RBFN,
123
Microsystem Technologies
Fig. 5 Comparison of
classification accuracy of non-
linear features
Table 5 MLPNN classifier’s performance for non-linear features Figure 7 compares the classification accuracy of
MLPNN, RBFN, LDA and QDA based on best performing
Features Accuracy (%) Sensitivity (%) Specificity (%)
linear feature (alpha power) non-linear feature (WE) and
RWE 86.67 86.67 86.67 combination of linear and non-linear feature (alpha
WE 88.33 90.00 86.67 power ? WE). It shows that combination of alpha power
and WE features works the best in all types of applied
classifier.
RWE 90.00 83.33 96.67 Highest classification accuracy of 93.33% was achieved by
WE 83.33 73.33 93.33 combination of alpha power and WE features in MLPNN
and RBFN classifiers. The improvement in accuracy by
combination of linear and non-linear feature is because of
the fact that the new feature set can analyse the signal’s
frequency and time domain behavior as well as the chaotic
Table 7 LDA classifier’s performance for non-linear features
and complex behavior of the signal.
Features Accuracy (%) Sensitivity (%) Specificity (%) Combination of two non-linear features, i.e., WE and
RWE 86.67 80.00 93.33 RWE also provided the highest classification accuracy of
WE 90.00 83.33 96.67 91.67% using LDA classifier which is higher than accuracy
obtained by any non-linear feature alone in any of the
applied classifier. The improvement in the accuracy is due
to the fact that when combination of two non-linear fea-
tures is used more amount of complex and chaotic behavior
Table 8 QDA classifier’s performance for non-linear features
of brain signal is captured.
Features Accuracy (%) Sensitivity (%) Specificity (%) Hosseinifarda et al. (2013) got similar result with
accuracy of 90% by combination of four non-linear fea-
RWE 80.00 86.67 73.33
tures (DFA, HFD, CD and Lyapunov exponent) using
WE 85.00 93.33 76.67
logistic regression classifier.
Alpha band power provided the highest accuracy among
the entire power band showing that MDD patients and
healthy individuals more significantly differ in alpha band
LDA, QDA respectively using combination of linear and than in any other power band. Similar results were found in
non-linear features. literature (Hosseinifarda et al. 2013; Mohammadi et al.
On the basis of above results it can be concluded that 2015).
combination of RWE and alpha power gives the best result Table 13, compares the present study with existing
in all kind of classifier. algorithm (Mumtaz et al. 2017b) on the published dataset.
123
Microsystem Technologies
Fig. 6 Comparison of
classification accuracy of
combination of linear and non-
linear features
123
Microsystem Technologies
Fig. 7 Comparison of
classification accuracy of
classifiers for the best
performing linear feature (alpha
power), non-linear feature (WE)
and combination of linear and
non-linear feature (alpha
power ? RWE)
Table 13 Comparative study with existing algorithm on the published data set
Paper name A wavelet-based technique to predict treatment outcome Detection of major depressive disorder using linear and non-
for major depressive disorder (Mumtaz et al. 2017a, b) linear features from EEG signal
Objective (i) Prediction of response of antidepressant’s treatment on Classification of MDD and healthy individuals based on
MDD patient using EEG signal EEG signal using different linear and non-linear features
(ii) Classification of MDD and healthy individuals based
on EEG signal
Features used EEG features computed with wavelet transform (WT) Linear features: band power and asymmetry
analysis, short-time Fourier transform (STFT) analysis, Nonlinear features: Relative Wavelet Energy (RWE),
and Empirical Mode decompositions (EMD) Wavelet Entropy(WE) and combination of linear and non-
linear features
Classifier used Logistic regression (LR) MLPNN, RBFN, LDA, QDA
Feature reduction (i) Rank-based feature selection method on the basis of Principal component analysis (PCA)
and selection receiver operating characteristic (ROC)
techniques used (ii) Minimum redundancy and maximum relevance
(mRMR) method
Results Highest classification accuracy was achieved with Highest classification accuracy was achieved with
combination wavelet, STFT and EMD features with combination alpha power and RWE in both MLPNN and
Rank Based selection technique RBFN using PCA for reduction
Accuracy: 90.5 ± 8.3% Accuracy: 93.33 ± 1.67%
Sensitivity: 91.6 ± 5.7% Sensitivity: 94.44 ± 3.68%
Specificity: 88.7 ± 7.5% Specificity: 87.78 ± 2.12%
https://figshare.com/articles/EEG-based_Diagnosis_and_Treatment_Outcome_Prediction_for_Major_Depressive_Disorder/3385168
Alpha asymmetry as well as theta asymmetry showed performs better than RBFN when the entire feature’s per-
good classification accuracy. A lot of study (Hinrikus et al. formance is compared.
2009; Ricardo-Garcel et al. 2010; Stewart et al. 2014; MLPNN gives good and reliable results due to its ability
Gollan et al. 2014) has been done in the field of alpha to represent complex non-linear behavior of the problem.
asymmetry which shows its potential power for high pre- This characteristic of MLPNN is mainly because of pres-
diction accuracy. ence of one or more hidden layers with differentiable
But this study reveals the potential power of high pre- activation function and the layers are highly connected.
diction accuracy of theta power which is an area of MLPNN being a non-parametric model can capture more
research still needed to be explored. minute details of the data which is helpful in predicting
In this study four classifiers: MLPNN, RBFN, LDA and future data.
QDA have been used. Both MLPNN and RBFN performed Future research work would concentrate on the specific
better than LDA and QDA in most of the cases. MLPNN regions of the brain which gets affected in depression. EEG
123
Microsystem Technologies
signal of more number of MDD patient’s needs to be Haykin S (2009) Multilayer perceptrons. In: Dworkin A, Mars D,
analyzed so that the results could be generalized. Opaluc W (eds) Neural networks and learning machines, 3rd
edn. Pearson Education, Cranbury, pp 1–263
Hinrikus H, Sukhova A, Bachmann M et al (2009) Electroencephalo-
graphic spectral asymmetry index for detection of depression.
5 Conclusion Med Biomed Eng Comput 47:1291–1299
Hosseinifarda B, Moradia MH, Rostami R (2013) Classifying
The study demonstrated that EEG signal can be effectively depression patients and normal subjects using machine learning
used in discriminating between MDD patients and healthy techniques and nonlinear features from EEG signal. Comput
Methods Programs Biomed 109:339–345
individuals. Combination of linear and non-linear feature James G (2013) Classification. In: Casella G, Fienberg S, Olkin I
or combination of non-linear features is also an effective (eds) An introduction to statistical learning with applications in
way of increasing the accuracy of the classifier. Along with R. Springer, New York, pp 138–150
alpha asymmetry, theta asymmetry can also be used for Jolliffe IT (2002) Principal component analysis, series: Springer
series in statistics, 2nd edn. Springer, New York, pp 1–147
diagnosis of depression. Among all the classifier, i.e., Joyce CA, Gorodnitsky IF, Kutas M (2003) Automatic removal of eye
MLPNN, RBFN, LDA and QDA, MLPNN outperformed movement and blink artifacts from EEG data using blind
all other classifiers in the given localized data set. component separation. In: Fabiani M, Jennings JR (eds)
Psychophysiology. Blackwell Publishing Inc, Malden,
pp 313–325
Jung TP, Humphries C, Lee TW et al (1998) Extended ICA removes
References artifacts from electroencephalographic recordings. Adv Neural
Inf Process Syst 10:1–7
American Psychiatric Association (1994) Diagnostic and statistical Jung TP, Makeig S, Humphries C et al (2000) Removing electroen-
manual of mental disorders, 4th edn. American Psychiatric cephalographic artifacts by blind source separation. Psychophys-
Association, Washington, DC, pp 339–345 iology. Cambridge University Press, Cambridge, pp 163–178
Bachmann M, Lass J, Suhhova A, Hinrikus H (2013) Spectral Mallat SG (1989) A theory for multi-resolution signal decomposition:
asymmetry and Higuchi’s fractal dimension measures of depres- the wavelet representation. IEEE Trans Pattern Anal Mach Intell
sion electroencephalogram. Comput Math Methods Med II:674–694
2013:1–9 Mohammadi M, Al-Azab F, Raahem B et al (2015) Data mining EEG
Bachmann M, Lass J, Hinrikus H (2017) Single channel EEG analysis signals in depression for their diagnostic value. BMC Med
for detection of depression. Biomed Signal Process Control Inform Decis Making 108:108–123
31:391–397 Mumtaz W, Xia L, Ali SSA et al (2017a) A wavelet-based technique
Bishop C (2006) Linear models for classification. In: Jordan M, to predict treatment outcome for major depressive disorder.
Kleinberg J, Scholkopf B (eds) Pattern recognition and machine PLoS One 2017:1–30
learning. Springer, Singapore, pp 186–189 Mumtaz W, Xia L, Ali SSA et al (2017b) Electroencephalogram
Bopardikar AS, Rao RM (1998) Wavelet transforms: Introduction to (EEG)-based computer-aided technique to diagnose major
Theory and Applications. Dorling Kindersley Publishing Inc, depressive disorder (MDD). Biomed Signal Process Control
New Delhi, pp 2–82 31:108–115
Bruder GE, Stewart JW, Hellerstein D et al (2012) Abnormal Puthankattil SD, Joseph PK (2012) Classification of EEG signals In
functional brain asymmetry in depression: evidence of biologic normal and depression conditions by ANN using RWE and
commonality between major depression and dysthymia. Psychi- signal entropy. J Mech Med Biol 12:1240019–1240032
atry Res 196:250–254 Puthankattil SD, Joseph PK (2014) Analysis of EEG signals using
Cusin C, Yang H, Yeung A et al (2009) Rating scales for depression. wavelet entropy and approximate entropy: a case study on
In: Baer L, Blais MA (eds) Handbook of clinical rating scales depression patients. Int J Bioeng Life Sci 8:420–424
and assessment in psychiatry and mental health. Current Clinical Ricardo-Garcel J, Gonzalez-Olvera JJ, Miranda E et al (2010) EEG
Psychiatry, Boston, pp 7–37 sources in a group of patients with major depressive disorders.
Delorme A, Makeig S (2004) EEGLAB: an open source toolbox for Int J Psychophysiol 71:70–74
analysis of single-trial EEG dynamics including independent Rodreguez-Bermudez G, Garcia-Laencina P (2015) Analysis of EEG
component analysis. J Neurosci Methods 134:9–21 signals using nonlinear dynamics and chaos: a review. Appl
Dharmadhikari AS, Tandle AL, Jaiswal SV et al (2018) Frontal theta Math Inf Sci 9:2309–2321
asymmetry as a biomarker of depression. East Asian Arch Rosso OA, Martin MT, Figliola A et al (2006) EEG analysis using
Psychiatry 28:17–22 wavelet-based information tools. J Neurosci Methods
Dien J (1998) Issues in the application of the average reference: 153:163–182
review, critiques and recommendations. Behav Res Methods Stewart JL, Coan JA, Towers DN et al (2014) Resting and task-
Instrum Comput 30:34–43 elicited prefrontal EEG alpha asymmetry in depression: support
Gandhi V (2014) Brain computer interfacing for assistive robotics. for the capability model. Psychophysiology 51:1–18
Electroencephalograms, recurrent quantum neural networks, and Tharwat A (2016) Linear vs. quadratic discriminant analysis classi-
user-centric graphical interfaces, 1st edn. Academic Press, fier: a tutorial. Int J Appl. Pattern Recognit 3(2):145–180
Cambridge, pp 21–29
Gollan JK, Hoxha D, Chihade D et al (2014) Frontal alpha EEG
asymmetry before and after behavioral activation treatment for Publisher’s Note Springer Nature remains neutral with regard to
depression. Biol Psychol 99:198–208 jurisdictional claims in published maps and institutional affiliations.
Grin-Yatsenko VA, Baas I, Ponomarev VA et al (2010) Independent
component approach to the analysis of EEG recordings at early
stages of depressive disorders. Clin Neurophysiol 281:281–289
123