EEG-based Subjects Identiﬁcation based on
Biometrics of Imagined Speech using EMD
Luis Alfredo Moctezuma and Marta Molinas
Department of Engineering Cybernetics, Norwegian University of Science and
Technology. Trondheim, Norway
Abstract. When brain activity is translated into commands for real ap-
plications, the potential for human capacities augmentation is promising.
In this paper, EMD is used to decompose EEG signals during Imagined
Speech in order to use it as a biometric marker for creating a Biometric
Recognition System. For each EEG channel, the most relevant Intrinsic
Mode Functions (IMFs) are decided based on the Minkowski distance,
and for each IMF 4 features are computed: Instantaneous and Teager
energy distribution and Higuchi and Petrosian Fractal Dimension. To
test the proposed method, a dataset with 20 subjects who imagined 30
repetitions of 5 words in Spanish, is used. Four classiﬁers are used for
this task - random forest, SVM, naive Bayes and k-NN - and their per-
formances are compared. The accuracy obtained (up to 0.92 using Linear
SVM) after 10-folds cross-validation suggest that the proposed method
based on EMD can be valuable for creating EEG-based biometrics of
imagined speech for Subjects identiﬁcation.
Keywords: Biometric security, Subjects identiﬁcation, Imagined Speech,
Electroencephalograms (EEG), Empirical Mode Decomposition (EMD)
Electroencephalography (EEG) is a popular non-invasive technique of Brain
Computer Interface (BCI), and it refers to the exploration of bioelectrical brain
activity registered during diﬀerent activation functions. EEG does not require
any type of surgery, however, compared with invasive techniques the signals ob-
tained are weaker. Another important term is the Electrophysiological source,
that refers to the neurological mechanisms adopted by a BCI user to stimulate
the brain signals .
Due to the easy setup and the little training required, this paper uses the
Electrophysiological source Imagined Speech , that refers to imagined or inter-
nal speech without uttering-sounds / articulating-gestures, to create a biometric
system for subjects identiﬁcation.
Due to the non-stationary and non-linear nature of brain signals, signal pro-
cessing tools like the Wavelet Transform [3,4,5] and power spectral density (PSD)
 capable of dealing with these properties, have been reported in the literature.
arXiv:1809.06697v1 [q-bio.NC] 13 Sep 2018
2 EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech
Most recent works have shown wavelets as a powerful tool to analyze brain sig-
nals. However, its main disadvantage is the need to ﬁt the best mother function
for the signal. This means that mother functions will be diﬀerent depending on
the task/neuro-paradigm/environment adopted.
Recently, the Empirical Mode Decomposition (EMD) has been employed to
analyze brain signals corresponding to diﬀerent tasks. It has shown to be robust
in decomposing non-stationary and non-linear time series, with the advantage
that it does not need a-priory deﬁnition of speciﬁc parameters to the signal, in
contrast with wavelet transform.
The interest in biometric recognition systems has increased in the las years,
since traditional security systems (security guards, smart cards, etc) poses serious
challenges of increased vulnerabilities. Current biometric security systems are
vulnerable due to a variety of attacks to skip the authentication process .
This is because authentication systems cannot discriminate between authorized
users and an intruder who fraudulently obtains the access privileges.
To tackle this problem, some researchers have explored the use of brain sig-
nals as a measure for a biometric security system. This is possible because any
human physiological and/or behavioral characteristic can be used as a biometric
feature as long as it satisﬁes the following requirements: universality, perma-
nence, collectability, performance, acceptability and circumvention. A biometric
recognition system is able to perform automatic Subjects identiﬁcation based
on their physiological and/or behavioral features with the advantage that a
single biometric trait can be used for the access into several accounts.
In that context, the main neuro-paradigms in the state-of-the-art are: senso-
rimotor activity, imagination of activities (Visual Counting and geometric ﬁgure
Rotation  and mental composition of letters , among others . One of the
challenges for subject identiﬁcation task is the feature extraction stage in order
to represent the brain signal captured from diﬀerent electrodes with a single vec-
tor, since it is impractical and computationally costly to use all data generated
by the brain.
Some authors report the use of imagined speech, for example  used EEG
signals from a small population of 6 subjects while imagining the syllables /ba/
and /ku/. The collected database consisted of 6 sessions and for each one 20
trials per subject from 128 channels with a sampling frequency of 1024 Hz. using
Electrical Geodesics device. For feature extraction they used the PSD for each
EEG signal, then autoregressive (AR) model coeﬃcients were computed for each
electrode using the Burg method . The classiﬁcation stage was performed
using the linear kernel of Support Vector Machine (SVM) classiﬁer and using
1-Nearest-Neighbor (k-NN). For these two syllables they obtained 99.76% and
99.41% of accuracy respectively. In the work presented in , resting-states were
used for subject identiﬁcation using Linear SVM. The dataset used consisted
of 40 subjects, and 192 instances per subject. The sampling frequency was 256
Hz with 64 channels. First, for pre-processing a band-pass ﬁlter (0.5-40 Hz) and
then the Common Average Reference were applied. For feature extraction the
Morlet Wavelet was used to extract power spectrum of 7 frequency bands, to
EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech 3
ﬁnally apply a downsampling to 32 Hz. The accuracies obtained in the best cases
were 100%, 96% and 72% respectively for 3 lengths of the signal (300, 60 and
30 seconds). However, in a real application, the registry of 300, 60 or even 30
seconds of a signal can be impractical and with high computational cost for real-
time. In addition the use of 128 or 64 channels does not support the portability
of the device.
As a ﬁrst step to create a robust method without a-priory deﬁnition of ad-
ditional parameters, a method based on EMD to extract features from brain
signals of Imagined Speech, is presented here.
1.1 Empirical Mode Decomposition
The EMD method is useful to decompose non-linear and non-stationary sig-
nals into a ﬁnite number of Intrinsic Mode Functions (IMFs) that satisﬁes two
1. The number of extrema and the number of zero crossings must be either
equal or diﬀer at most by one.
2. At any point, the mean value of the envelope deﬁned by the local maxima
and the envelope deﬁned by the local minima is zero.
The method decomposes a signal into oscillatory components by applying a
process called sifting. The sifting process for the signal x(t) can be summarized
as shown in the algorithm 1:
Data: Time serie = x(t)
sifting = True;
while sifting = True do
1. Identify all upper extrema in x(t)
2. Interpolate the local maxima to form an upper envelope u(x).
3. Identify all lower extrema of x(t)
4. Interpolate the local minima to form an lower envelope l(x)
5. Calculate the mean envelope:
m(t) = u(x)+l(x)
6. Extract the mean from the signal:
h(t) = x(t)−m(t)
if h(t)satisﬁes the two IMF conditions then
h(t) is an IMF;
sifting = False ; Stop sifting
sifting = True ; Keep sifting
if x(t)is not monotonic then
Algorithm 1: The sifting process for a signal x(t)
1.2 IMFs selection
The EMD is a powerful tool to decompose a non-stationary signal, however some
IMFs that contain limited information may appear in the decomposition because
4 EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech
the numerical procedure is susceptible to errors . To select the IMFs that
contain the most relevant information about the signal, the methods presented
in [14,15] were applied and compared, to ﬁnally use the method proposed in 
that employs the Minkowski Distance (dmink), as follow.
dmink = n
where xiand yiare the i-th respective samples of the observed signal and
the extracted IMF.
According to the authors, the redundant IMFs have a shape and frequency
content diﬀerent than those of the original signal, which means that when a IMF
is not appropriate, the dmink presents a maximum value.
In this work, a new method for feature extraction based on EMD is proposed.
In the next section the method is described in brief. Then, the application of the
proposed method for Subjects identiﬁcation is explained.
2 Description of the method
The main contribution of the proposed method is the feature extraction stage,
that consists on applying the Empirical Mode Decomposition (EMD) to obtain
5 Intrinsic Mode Functions (IMF) per channel of the EEG data.
To select the most relevant IMFs, the Minkowski Distance was computed .
Once the most relevant IMFs from all instances in the dataset were obtained,
it turned out that the number of IMFs was diﬀerent depending of the size of
the signal and the imagined word. However, to obtain meaningful features it is
necessary to have the same number of IMFs from all instances. To cope with this,
the IMFs selected were limited to the minimum relevant IMFs in all instances,
which in this case were only the ﬁrst 2 IMFs.
Then for each IMF, 4 features were computed: Instantaneous energy, Teager
energy, Higuchi fractal dimension and Petrosian fractal dimension, as it is
shown in the ﬁgure 1. All features per IMF and per channel were concatenated
to obtain a feature vector per instance. Once the feature vectors were obtained,
they were used to train 4 machine learning-based classiﬁers (random forest, naive
Bayes, Support Vector Machine (SVM) and K-Nearest Neighbors (k-NN)) in
order to compare their performances.
In the following, the step-by-step procedure proposed in this work to identify
Subjects by using the EEG of their imagined speech, is described.
2.1 Feature extraction
When the most relevant IMFs were selected, 4 features were computed for each
one in order to reduce the feature vector to obtain a good representation of the
signal. The ﬁrst 2 features used are related with the energy distribution and
the others two with Fractal dimensions. Each feature tested in this work is here
described in brief.
EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech 5
Fig. 1. Flowchart summarizing the feature extraction stage.
–Instantaneous Energy: gives the energy distribution in each band :
–Teager Energy: This energy operator reﬂects variations in both amplitude
and frequency of the signal and it is a robust parameter as it attenuates
auditory noise [16,17].
r=1 (wj(r))2−wj(r−1) ∗wj(r+ 1)!(3)
–Higuchi Fractal Dimension: The algorithm approximates the mean length
of the curve using segments of ksamples and estimates the dimension of a
time-varying signal directly in the time domain . Considered a ﬁnite set
of observations taken at a regular interval: X(1), X(2), X (3), .., X(N). From
this series, a new one Xm
kmust be constructed,
k:X(m), X(m+k), X(m+ 2k), .., X m+N−m
Where m= 1,2, .., k,mindicate the initial time and kthe interval time.
Then, the length of the curve associated to each time series Xm
computed as follow:
Lm(k) = 1
i=1 X(m+ik)−Xm+ (i−1)k! N−1
6 EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech
Higuchi takes the mean length of the curve for each k, as the average value
of Lm(k), for m= 1,2, ..., k and k= 1,2, ..., kmax, that it is calculated as:
L(k) = 1
–Petrosian Fractal Dimension: can be used to provide a fast compu-
tation of the fractal dimension of a signal by translating the series into a
binary sequence .
F DP etrosian =log10 n
log10 n+ log10 n
Where nis the length of the sequence and N∇is the number of sign changes
in the binary sequence.
2.2 Classiﬁers and validation
At this point, the features vector have the same features per each instance with
an assigned tag. This allows the use of machine learning methods. In this work,
the machine learning methods random forest, naive Bayes, SVM and k-NN were
used. For SVM, all experiments were reproduced with the kernels Linear, RBF
(Radial Basis Function) and Sigmoid. In the random forest case, the experiments
were reproduced with diﬀerent tree depths (2, 3, 4, 5) using the Gini impurity.
k-NN was tested with diﬀerent number of neighbors (1, 2, 3, 4, 5, 6, 7, 8, 9).
The accuracy with the 4 classiﬁers was estimated to evaluate their perfor-
mances using 10-folds cross-validation.
3 Dataset and experiments
In this section, the dataset used to test the proposed method in two diﬀerent
experiments for Subjects identiﬁcation is described in brief .
The purpose of the ﬁrst experiment is to show that the Subject can be iden-
tiﬁed regardless of the imagined word, to show if a biometric system using diﬀer-
ent password per Subject can be possible. The second experiment aim is to ﬁnd
whether the accuracy is higher if the password is pre-deﬁned. In others words,
using the same imagined word as biometric security measure.
The dataset consists of EEG signals from 27 subjects recorded using EMOTIV
EPOC device while imagining 33 repetitions of ﬁve imagined words in Spanish;
arriba, abajo, izquierda, derecha and seleccion, corresponding to the English
words up, down, left, right and select.
EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech 7
Each repetition of the imagined words was separated by a state of rest. The
protocol for EEG signal acquisition is described in details in .
EEG signals were recorded from 14 channels which were placed on the head
according to the 10-20 international system , with a sample frequency of 128
For the next experiments the ﬁrst 20 subjects and the ﬁrst 30 repetitions per
each of the 5 imagined words were used . In summary, the terms used along the
paper are the following.
–S∇= 20: Subjects.
–W∇= 5: The imagined words in the dataset.
–R∇= 30: Repetitions per imagined word.
–C∇= 14: Channels used in all instances.
–I MF s∇= 2: IMF per channel.
–F∇= 4: Features per IMF, corresponding to Teager energy, Instantaneous
energy, Higuchi Fractal Dimension and Petrosian Fractal Dimension.
According to the proposed method the feature vector size per subject is
F∇∗I MF s∇∗C∇= 112. Next, the speciﬁc setup for the experiments and the
results are presented.
3.3 Subject level analysis
In this experiment, 4 classiﬁers were used in order to compare their performances
and each classiﬁer has S∇classes, and per class R∇∗W∇= 150 instances.
In the table 1 the accuracies obtained with the proposed method are shown.
Table 1. Accuracy obtained when all imagined words were joined for subjects identi-
random forest 0.64
naive Bayes 0.68
The aim of this experiment is to show that the method can be used for
subjects identiﬁcation with high accuracy rates. According to the results in the
table 1, the classiﬁer SVM was the best. As it was mentioned before, SVM was
tested with diﬀerent kernels and for this experiment the best one was the Linear
8 EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech
3.4 Word level analysis
In this experiment, the classiﬁers were trained using all words separately, in
order to explore if the proposed method works best for a speciﬁc word. Each
classiﬁer has S∇classes, and per class R∇instances.
Table 2. Accuracy obtained per imagined word for subjects identiﬁcation
Classiﬁer Up Down Left Right Select
random forest 0.78 0.77 0.73 0.73 0.75
SVM 0.91 0.87 0.88 0.84 0.92
naive Bayes 0.90 0.85 0.88 0.85 0.89
k-NN 0.85 0.80 0.81 0.79 0.88
In this experiment also the highest accuracy was obtained when using the
SVM classiﬁer with the Linear kernel, obtaining the accuracy of 0.92. On the
other hand, the lowest performance was obtained using the classiﬁer random
4 Discussion and Conclusions
In this paper, a method based on EMD for Subjects identiﬁcation from EEG
signals of imagined speech was presented. The proposed method was applied to
a dataset of Imagine Speech with encouraging results. The accuracies obtained
suggest that the the use of imagined speech for Subjects identiﬁcation, specially
using the classiﬁer Linear SVM, can be eﬀective and worth exploring further.
When the imagined words were joined to observe if it is possible to identify
a Subject regardless of the imagined word, the highest accuracy obtained was
0.84 using Linear SVM. Then, when the second experiment was carried out
to observe the accuracies using the imagined words separately, the maximum
accuracies reached were also using Linear SVM : 0.91, 0.87, 0.88, 0.84 and 0.92
respectively per each imagined word.
In the work presented in  the Common Average Reference (CAR)
was used to improve the signal-to-noise ratio. Then, the feature extraction was
based on Instantaneous and Teager energy distribution of 4 decomposition levels
of Wavelet using the mother function Biorthogonal 2.2, and random forest for
classiﬁcation. The accuracies obtained using the imagined word select were 0.96
and 0.93 respectively. In this paper, the highest accuracy (Using Linear SVM)
obtained for the imagined word select was 0.92, which is slightly lower than the
above ones. However, with the inherent adaptivity of EMD for feature extraction,
there is no need to pre-deﬁne any mother function for the particular task or
neuro-paradigm. In addition, the EMD inherently improves the signal-to-noise
ratio by removing the noise in the ﬁrst IMFs.
EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech 9
A limitation of the proposed method is the use of a dataset of brain signals
from only 20 subjects. In future, it will be necessary to reproduce the experiments
using a larger population in order produce an alternative competitive system to
the current biometric security systems used in industry. Further eﬀorts will be
made to explore alternative techniques for IMFs selection and for EEG channels
selection, since it is well known that speciﬁc channels will provide more relevant
information than others for the distinct task selected for subject identiﬁcation.
Acknowledgments. This work was supported by Enabling Technologies -
NTNU, under the project “David versus Goliath: single-channel EEG unrav-
els its power through adaptive signal analysis - FlexEEG”.
1. Bashashati, Ali, Mehrdad Fatourechi, Rabab K. Ward, and Gary E. Birch: A survey
of signal processing algorithms in braincomputer interfaces based on electrical brain
signals. Journal of Neural engineering 4, no. 2 (2007): R32.
2. Desain, Peter, Jason Farquhar, Pim Haselager, Christian Hesse, and R. S. Schaefer:
What BCI research needs. In Proc. ACM CHI 2008 Conf. on Human Factors in
Computing Systems (Venice, Italy) (2008):.
3. Moctezuma, Luis Alfredo, Maya Carrillo, Luis Villase˜nor Pineda, and Alejandro A.
Torres Garc´ıa: Hacia la clasiﬁcaci´on de actividad e inactividad ling¨uistica a partir
de senales de electroencefalogramas (EEG). Research in Computing Science 140
4. Torres-Garc´ıa, A. A., C. A. Reyes-Garc´ıa, L. Villase˜nor-Pineda, and J. M. Ram´ırez-
Cort´ıs: An´alisis de se˜nales electroencefalogr´aﬁcas para la clasiﬁcacin de habla imag-
inada. Revista mexicana de ingenier´ıa biom´edica 34, no. 1 (2013): 23-39.
5. Nishimoto, Takashi, Yoshiki Azuma, Hiroshi Morioka, and Shin Ishii: Individual
Identiﬁcation by Resting-State EEG Using Common Dictionary Learning. In Inter-
national Conference on Artiﬁcial Neural Networks. Springer, Cham (2017): 199-207.
6. Brigham, Katharine, and BVK Vijaya Kumar: Subject identiﬁcation from electroen-
cephalogram (EEG) signals during imagined speech. In Biometrics: Theory Applica-
tions and Systems (BTAS), 2010 Fourth IEEE International Conference on (2010):
7. Jain, Anil K., Arun Ross, and Umut Uludag: Biometric template security: Chal-
lenges and solutions. In Signal Processing Conference, 2005 13th European (2005):
8. Ashby, Corey, Amit Bhatia, Francesco Tenore, and Jacob Vogelstein: Low-cost elec-
troencephalogram (EEG) based authentication. In Neural Engineering (NER), 2011
5th International IEEE/EMBS Conference on (2011): 442-445.
9. Palaniappan, Ramaswamy: Electroencephalogram signals from imagined activities:
A novel biometric identiﬁer for a small population. In International Conference on
Intelligent Data Engineering and Automated Learning. Springer, Berlin, Heidelberg
10. Del Pozo-Banos, Marcos, Jes´us B. Alonso, Jaime R. Ticay-Rivas, and Carlos M.
Travieso: Electroencephalogram subject identiﬁcation: A review. Expert Systems
with Applications 41, no. 15 (2014): 6537-6554.
10 EEG-based Subjects Identiﬁcation based on Biometrics of Imagined Speech
11. Steven, M. Kay: Modern spectral estimation: theory and application. Signal Pro-
cessing Series (1988):.
12. Huang, Norden E., Zheng Shen, Steven R. Long, Manli C. Wu, Hsing H. Shih,
Quanan Zheng, Nai-Chyuan Yen, Chi Chao Tung, and Henry H. Liu: The empirical
mode decomposition and the Hilbert spectrum for nonlinear and non-stationary
time series analysis. In Proceedings of the Royal Society of London A: mathematical,
physical and engineering sciences, vol. 454, no. 1971 (1998): 903-995.
13. Rilling, Gabriel, Patrick Flandrin, and Paulo Goncalves: On empirical mode de-
composition and its algorithms. In IEEE-EURASIP workshop on nonlinear signal
and image processing, vol. 3 NSIP-03, Grado (I), (2003): 8-11.
14. de Souza, Douglas Baptista, Jocelyn Chanussot, and Anne-Catherine Favre: On
selecting relevant intrinsic mode functions in empirical mode decomposition: An
energy-based approach. In Acoustics, Speech and Signal Processing (ICASSP), 2014
IEEE International Conference on (2014): 325-329.
15. Boutana, Daoud, Messaoud Benidir, and Braham Barkat: On the selection of in-
trinsic mode function in EMD method: application on heart sound signal. In Applied
Sciences in Biomedical and Communication Technologies (ISABEL), 2010 3rd In-
ternational Symposium on (2010): 1-5.
16. Didiot, Emmanuel, Irina Illina, Dominique Fohr, and Odile Mella: A wavelet-based
parameterization for speechmusic discrimination. Computer Speech & Language 24,
no. 2 (2010): 341-357.
17. Jabloun, Firas, and A. Enis Cetin: The Teager energy based feature parameters for
robust speech recognition in car noise. In Acoustics, Speech, and Signal Processing.
1999 IEEE International Conference on, vol. 1. (1999): 273-276.
18. T. Higuchi: Approach to an irregular time series on the basis of the fractal theory.
Physica D, Nonlinear Phenomena, vol. 31, (1988): 277-283.
19. A. Petrosian: Kolmogorov complexity of ﬁnite sequences and recognition of diﬀer-
ent preictal eeg patterns. In Computer-Based Medical Systems, Proceedings of the
Eighth IEEE Symposium on (1995): 212-217.
20. Jasper, Herbert: Report of the committee on methods of clinical examination in
electroencephalography. Electroencephalogr Clin Neurophysiol 10 (1958): 370-375.
21. Moctezuma, Luis Alfredo, Marta Molinas, A. A. Torres Garc´ıa, Luis Villase˜nor
Pineda and Maya Carrillo: Towards an API for EEG-Based Imagined Speech clas-
siﬁcation. International conference on Time Series and Forecasting (2018):.
22. Bertrand, O., F. Perrin, and J. Pernier: A theoretical justiﬁcation of the aver-
age reference in topographic evoked potential studies. Electroencephalography and
Clinical Neurophysiology/Evoked Potentials Section 62, no. 6 (1985): 462-464.