UNIT 3

Feature extraction
❖Extraction of useful information from the highly
complex EEG data.
❖It is achieved by applying suitable feature extraction
on EEG signals which are acquired by subjects
during performing a specific mental activity.
❖These features are fed to classifier for the training,
where it gets learning to identify the pattern class.
❖The EEG signals are highly nonstationary and
dynamically changes due to technical and
biological effects such as subject attention, the
sessions variability, mental state, anatomical
differences among subjects, amplifier and
ambient noise.
❖Due to these reasons, it is difficult to classify BCI
patterns accurately.
❖Therefore to enhance the performance of EEG
based BCI systems, the selection of appropriate
feature extraction technique is a very important
issue.
Feature extraction methods
Temporal features i.e time domain features
represents employing the EEG signal values at
distinct time windows or at distinct time points.
Frequency domain features are also called spectral
features which represent the signal power in the
specific frequency band.
However, for the EEG signals having nonstationary in
nature, time-frequency methods are useful, which
can provide useful information by taking into
consideration the dynamic changes.
Spatial features deal with the spatial representation
of the signal i.e the selection of most appropriate
channels selection for the specific task.
A.)Time Domain Features
These features measure the temporal variations
within time-locked EEG signal amplitudes.
1) Hjorth Parameters:
These provide a fast way of computing three important
characteristics of a time varying signal, namely, Activity,
Mobility, and Complexity.

❑ Activity parameter of the signal z(t) is computed

as in equation 1, which is the variance of the time
function, which designates the power spectra
surface in the frequency domain.
❑Mobility parameter represents standard deviation
proportion or mean frequency of power spectrum.
Mobility is computed as in equation 2, where z’(t)
represents the first derivative of the signal z(t).

❑Complexity parameter computed as in equation 3,

which indicates the signal shape similarity to a pure
sine wave, where value converges to 1 as the
similarity index is high.
 2) Statistical Features:
Various statistical measures characterize the EEG
time series. The following six features are widely
used in BCI studies and computed by considering N
i.e the number of data samples and the signal x(i).

❑ Energy: The energy of signal is calculated as in

equation 4.

❑ Entropy: Entropy computes randomness in the

signal and computed as in equation 5.
❑Mean: The mean (µ) which is average of the
signal calculated as shown in equation 6

❑Std Deviation: The standard deviation measures

how the data {x1, x2, ..., xN } are spread out. The
standard deviation is computed as in equation 7.
❑Skewness: The asymmetry of the data samples
around mean is calculated with skewness as
shown in equation 8.

where µ3 is the 3rd order moment which is

calculated as shown in equation 9.
❑Kurtosis: Kurtosis measure the ‘tailedness’ and ‘outlier’
characteristic of the distribution of data. Kurtosis is
computed as shown in equation 10.

where µ4 is the 4th order moment which is calculated as

shown in equation 11.
 3) Fractal Dimension:
▪The Fractal Dimension (FD) is a statistical index
which represents the self-similarity measure of
signal over some space or time interval.
▪EEG has a fractal nature, hence fractal pieces
can be used to obtain features.
▪Since the complexity and limited predictability of
EEG signals, Fractal dimension measure
provides a ‘complexity’ of the time-varying brain
signal.
▪Several classical methods have been proposed
to compute the FD value such as apparent
entropy, Kolmogorov-Sinai entropy and
Correlation dimension.
▪However, the evaluation of these methods is time
consuming.
 4) Kalman Filter:
▪Uncertainty representation is important in BCI
because potentially disastrous actions based on
poor estimates can be avoided if the amount of
uncertainty associated with an estimate is taken
into account before committing to a decision.
▪Bayesian filtering techniques provide a statistically
sound methodology for estimating signal
properties and their uncertainty.
▪The Kalman filter is perhaps the best known of
Bayesian filtering algorithms.
▪The time-varying coefficients can be updated
online using a recursive least square optimization
procedure such as Kalman filtering.
▪The coefficients capture the local statistical
structure of the signal as it evolves over time and
can be used as features in further processing e.g
classification in a BCI.
▪Kalman filters assume that the dynamics and
measurement process are linear and Gaussian.
▪This simplifying hypothesis may not dominance
true in many real-world examples.
 5) Particle Filter:
▪EEG neural signals acquired from human scalp
are nonlinear in nature.
▪However, the various linear regression model are
unable to reflect the nonlinear component of
EEG.
▪To overcome this drawback nonlinear decoding
model i.e particle filter can be used.
▪A particle filter is a method of estimating a
posterior distribution over the hidden state for
non-linear non-Gaussian processes.
▪It is achieved by applying Monte Carlo
simulations based recursive Bayesian filter.
▪The particle filtering is easy to implement and
scales very well i.e embarrassingly parallel.
▪Nevertheless, the shortcoming of particle filtering
is computationally expensive since a good
particle filter requires a lot of particles.
▪If the distribution is unimodal, it is good practice
to use a Kalman filter.
▪Particle filters are nondeterministic i.e they can
produce different outputs for the same input,
which make them difficult to predict and debug.
6) State-of-the-art:
Though the time domain features are not dominant in
EEG BCI, the aforementioned methods are useful to
identify the time series analysis of EEG signals.
In general the Horth parameters were widely used in
motor imagery EEG studies.
Hu et al. proposed the investigation of the learners
affect during the learning process by applying the
horth parameters in conjunction with Autoregressive
models (AR) and nonlinear features including
Singular-value Decomposition Entropy (SVDen),
Approximate Entropy (ApEn), the correlation
dimension (D2), the largest Lyapunov exponent, the
Kolmogorov entropy, the spectral entropy and
C0-complex.
The statistical features are widely used in conjunction
with other feature extraction techniques like dwt [12]–[14].
Several Fractal dimension computation methods such as
Fractal Brownian Motion [15], Sevciks method [16],
Higuchi algorithm [17] or Boxcounting [18] were
employed.
In [19] the fractal dimension coupled with dwt is used for
EEG signal classification.
Fractal dimensions are also employed in seizure
detection [104] and classifying depression patients [105].
Luke and Wouters [20] presented Auditory steady-state
responses (ASSR) using Kalman filter analysis and
illustrate several benefits over DFT methods.
Moreover, the Kalman filters were used in Epilepsy
patients EEG studies [21], [22]. However, the particle
filters were employed in P300-BCI paradigm [23], [24].
B. Frequency Domain Analysis
Noninvasive approaches such as EEG are based
on signals that reflect the activity of several
thousands of neurons.
The recorded signal thus is able to capture only
the correlated activities of large populations of
neurons, such as oscillatory activity.
For example, overt and imagined movements
typically activate premotor and primary
sensorimotor areas, resulting in amplitude/power
changes in the mu (8–12 Hz), beta (13–30 Hz)
and gamma (>30 Hz) rhythms in EEG/ECoG.
The existence of such oscillatory activity makes
analysis, such as Fourier analysis, of the signals
in the frequency domain particularly useful.
1) Fourier Analysis
▪The basic idea behind Fourier analysis is to
decompose a signal into a weighted sum of sine
and cosine waves of different frequencies.
▪Suppose you are given a step function that is a
constant positive value for some time and then
becomes a constant negative value, followed by
the original positive value again.
▪The step function can be approximated by adding
sine waves of different frequencies, each
weighted by a different coefficient (amplitude).
▪The step function can thus be decomposed into a
set of sine functions (a potentially infinite number
of them) of specific frequencies and amplitudes.
▪Fourier analysis involves decomposing a time-
varying signal s ( t ), defined over an interval t = - T /2
to t = T /2, into a weighted sum of sine and cosine
waves of different frequencies:

where ω is the angular frequency , defined as ω = 2

π / T , and f is the ordinary frequency (measured in
Hertz or cycles per second), defined as f = 1/ T .
The time interval T can be viewed as the period of a
periodic signal s ( t ).
The above expansion of s(t) into a sum of infinite
terms is called a Fourier series or Fourier expansion.
The Fourier series for a signal s( t ) then becomes:
The transformation to the set of coefficients cn given by
Equation 4.6 is also called the Fourier transform (FT) of
the signal s ( t ).
The transform is reversible: the original signal can be
recovered given the coefficients cn using Equation 4.5 –
this is called the inverse Fourier transform (IFT).
2 Discrete Fourier Transform (DFT)
▪For BCI applications, the brain signals are typically
sampled at discrete time intervals. The discrete
Fourier transform (DFT) takes as input a time series
S(t) sampled at time points t = 0, …, T - 1 and
transforms it to the corresponding complex Fourier
coefficients :

where ω = 2 π / T as before. The inverse discrete

Fourier transform (IDFT) is similarly defined as:
▪The complex Fourier coefficient C(n) captures both the
amplitude and phase of the nth sinusoidal component.
These can be recovered using the polar form of complex
numbers as:

where Re( x ) and Im( x ) denote the real and imaginary

parts of x .
▪The amplitude values A(n ) for n = 0,…, T - 1 define the
amplitude spectrum of the signal while the φ ( n ) values
define the phase spectrum. While the amplitude spectrum
can be used for this purpose, it is more common to
square the amplitude values and use the power spectrum
of the signal:
3) Fast Fourier Transform (FFT)
▪One can compute the DFT based on its definition
above, but for a signal with T points, this takes
approximately T2 arithmetical operations. The
running time of the algorithm is thus quadratic in
signal size T .
▪For very large T (e.g., in the millions), this can be
quite slow.
▪The fast Fourier transform (FFT) is a more
efficient way of computing the DFT. It runs in time
approximately T log T , which can result in huge
savings in computation time for large sizes T .
▪The most common FFT algorithm, the Cooley-
Tukey algorithm , uses a “divide and conquer”
strategy and recursively breaks down a DFT into
many smaller DFTs.
▪Most signal- processing packages come with an
FFT implementation, making the FFT the most
commonly used method for transforming a time-
varying signal to the frequency domain.
4) Spectral Features
▪Many BCI systems rely on features extracted
from the power spectrum of a brain signal such
as EEG or ECoG over a time interval.
▪The power spectrum is first computed using an
FFT, and the power in a particular frequency
band is used as a spectral feature in further
analysis (such as classification).
▪For example, given that motor movement or
imagery is known to reduce the power in the mu
frequency band (8–12 Hz), we could use the
power in the mu band as a feature in a BCI to
allow a subject to move a cursor using motor
imagery.
▪Another common approach is to use motor
screening to find subject- specific frequency
bands: the subject is asked to per form a variety
of movements, and the frequency bands that
exhibit robust changes in power during
movement are then utilized in subsequent BCI
experiments involving movement imagery.
▪A more sophisticated approach is to utilize a
bank of spectral features and allow a machine
learning algorithm to automatically select
features that enhance classification accuracy on
test data.
C. Time-Frequency Domain
Features
In general Time-Frequency representations of the
signal are the most complete and widely used
methods for non stationary EEG based BCI studies.

1) Matched Filtering:
▪The Matched Filtering (MF) is a feature extraction
technique which can detect the specified pattern from
the unknown EEG signals based on its match with
the templates, where templates are the known
signals.
▪The correlation between templates and unknown
EEG signals represents user's intention.
▪The better correlation intends higher matching between
intention of the user and template.
▪In general, as shown in equation 12 every matched filter
easily modelled as addition of sinusoidal components
which are the harmonically related.

where n denotes the template sample number, N is the total

number of harmonics to model, the sampling frequency fs and
fundamental frequency of template fF.

▪The FFT spectrum is used to derive ak and φk which

represents the amplitude and phase of individual
harmonics respectively.
 2) Short Time Fourier Transform:
▪The Fourier transform represents an original
signal with basis functions namely, sines and
cosines of different frequencies.
▪However, because sines and cosines occupy an
infinite temporal extent, the Fourier transform
does a poor job of representing signals that are
finite and non-periodic or having sharp peaks and
discontinuities.
▪However, brain signals such as EEG are typically
non-stationary (i.e statistical properties vary over
time), breaking the assumption of a stationary
signal in Fourier analysis.
▪One solution is to perform Fourier analysis over
short-time windows, a procedure known as
short-term Fourier transform (STFT).
▪The STFT deals with the problem of window size,
where small windows providing good temporal
resolution however poor frequency resolution,
while large windows providing better frequency
resolution yet poor temporal resolution.
▪This realization leads to wavelet transform which
achieves the excellent trade-off among temporal
and frequency resolution.
3) Wavelet:
The wavelet transform (WT) utilizes finite basis
functions called wavelets, such are translated
and scaled copies of a single finite length
waveform known as mother wavelet.
By using basis functions at different scales, the
wavelet transform allows a signal to be analysed
at multiple resolutions which allow representing
signals that are non periodic or have sharp
discontinuities.
The wavelets are a most suitable and powerful
tool for the transient EEG signal analysis. The
Continuous Wavelet Transform (CWT) of signal
x(t) with wavelet function as in equation 15.
where W(ST) is the wavelet coefficient of the
signal having frequency with scale S and time T
of the wavelet function S T (t). The symbol *
denotes the complex conjugate. The wavelet
function S T (t) is a dilated and shifted version of
mother wavelet (t).

A mother wavelet has ability to make multiple

shapes and satisfies the condition in equation 17.
Moreover, The CWT is like template matching or
matched filter where the cross variance between
the signals is calculated.
Wavelet template is always preferable than
classical template matching due to its special
properties.
However, CWT is more complex and redundant
since it involves the signal analysis with multiple
dilations and shifting to mother wavelet at a high
number of frequencies.
The Discrete Wavelet Transform (DWT)
translates and dilates mother wavelet at discrete
values and hence overcome the short coming of
CWT.
Though the DWT is more popular in BCI studies,
CWT is still applied in EEG based BCI research.
One of the reason may be CWT provides subtle
information where dwt inadequate to extract such
information.