COMPRESSED SENSING FRAMEWORK FOR EEG COMPRESSION
Selin Aviyente
Department of Electrical and Computer Engineering,
Michigan State University, East Lansing, MI, 48824
ABSTRACT
Many applications in signal processing require the efficient
representation and processing of data. The traditional ap-
proach to efficient signal representation is compression. In
recent years, there has been a new approach to compression
at the sensing level. Compressed sensing (CS) is an emerging
field which is based on the revelation that a small collection
of linear projections of a sparse signal contains enough in-
formation for reconstruction. In this paper, we propose an
application of compressed sensing in the field of biomedical
signal processing, particularly electroencophelogram (EEG)
collection and storage. A compressed sensing framework is
introduced for efficient representation of multichannel, mul-
tiple trial EEG data. The proposed framework is based on
the revelation that EEG signals are sparse in a Gabor frame.
The sparsity of EEG signals in a Gabor frame is utilized for
compressed sensing of these signals. A simultaneous orthog-
onal matching pursuit algorithm is shown to be effective in
the joint recovery of the original multiple trail EEG signals
from a small number of projections.
Index Terms— Biomedical signal processing, Electroence-
phalography, Signal Reconstruction, Sampling
1. INTRODUCTION
Electroencephalogram (EEG) is a commonly used brain imag-
ing method with high temporal resolution with applications
in neurology and psychology [?]. In a typical clinical set-
ting, EEG signals are collected over multiple channels, rang-
ing from 64 to 128 channels, and with multiple trials, i.e. the
same stimulus may be repeated multiple times to obtain high
signal-to-noise ratio estimates of the evoked response. One of
the biggest challenges with EEG data collection and analysis
is the amount of data that needs to be stored and processed.
EEG activity collected from a single subject can correspond to
2-3 hours of data. Therefore, there is a growing need for data
reduction or compression methods that can efficiently com-
press the EEG data into a few number of samples. An effi-
cient compression of the EEG data would reduce the amount
of data that needs to be stored and would also allow for the
fast wireless transmission of the collected EEG data in a clin-
ical setting.
1-4244-1198-X/07/$25.00 ©2007 IEEE 181
In recent years, a new paradigm, named Compressed Sens-
ing (CS), has been proposed for signal sensing and compres-
sion [1, 2]. CS builds on the revelation that a signal having
a sparse representation in one basis can be recovered from a
small number of projections onto a second basis that is inco-
herent with the first. This revelation has promising implica-
tions for applications for signal acquisition and compression.
With no a priori knowledge of a signal’s structure, a sensor
node could simultaneously acquire and compress the signa,
preserving the critical information. Some recent applications
of this theory include the single pixel camera [3] and com-
pressed sensing for rapid MR imaging [4].
In this paper, we use the framework of CS for EEG sig-
nal compression and reconstruction. We first show that EEG
signals are sparse in a Gabor frame based on an empirical
study using a large set of EEG data. Next, we apply the CS
framework to compress single-trial EEG recordings using a
few number of projections onto an i.i.d. Gaussian basis. The
reconstruction of the actual EEG signal from these projections
is achieved using orthogonal matching pursuit algorithm as
proposed in [5]. After the sparsity of individual EEG signals
is established, we extend the CS framework to joint recov-
ery of multiple signals using recent results in distributed com-
pressed sensing [6]. The joint sparsity of multiple EEG sig-
nals is shown and a simultaneous orthogonal matching pursuit
algorithm is used to reconstruct multiple recordings simulta-
neously.
2. BACKGROUND ON COMPRESSED SENSING
In this section, we will give a brief background on compressed
sensing theory. Given a signal x and a dictionary Ψ, x is
said to be sparse if x can be well approximated by a lin-
ear combination of a small set of vectors from Ψ, i.e. x ≈
 K
i=1 ani ψni , where K << N . The CS theory states that
it is possible to construct an M × N measurement matrix Φ,
where M << N , and reconstruct x from the measurements
y = Φx [1]. It has been shown that the signal can be recov-
ered from its measurements when the measurement matrix is
incoherent with the dictionary that the signal is sparse over.
Some choices for the measurement matrix, Φ, are a random
matrix with i.i.d. Gaussian entries or the Bernoulli (±1) ma-
trix. Using such a matrix Φ of size cK × N , where c is an
SSP 2007
oversampling factor, it was shown that it is possible, with high 12
x 10
−3

25
probability, to recover any signal that is K-sparse in the dic- Actual Signal
Reconstructed Signal
tionary Ψ from its projection onto Φ. 11 20

Several algorithms have been proposed for recovering x 15

Normalized Mean Square Error

from the measurements y by solving the following optimiza- 10

tion problem: 9
10

Amplitude
5
argmina a1 subject to ΦΨa = y (1) 8
0

The canonical approach uses linear programming, but has high 7

−5
computational complexity. Greedy algorithms such the match-
ing pursuit (MP) and orthogonal matching pursuit (OMP) re- 6 −10

quire fewer computations at the expense of slightly more mea- 5 −15

surements [5]. 0 5 10
Number of Dictionary Elements
15 20 0 100 200 300 400 500
Time (ms)
600 700 800

(a) (b)

3. SPARSITY OF EEG SIGNALS
In previous work, it has been shown that EEG signals can be
efficiently parameterized in terms of the elements of a Ga-
bor dictionary [7, 8, 9]. The Gabor dictionary has been com-
monly used to extract different time-frequency features from
the EEG signals for event characterization and classification
as in [8]. However, the sparsity of these signals have not been
officially shown on large sets of data.
Given a Gabor dictionary with elements,
1 (n − n0 )2
gi (n; n0 , f0 , σ) = √ exp(− ) exp(−j2πf0 n),
2πσ 2σ 2
(2)
where n0 and f0 are the time and frequency centers of the
Gabor logon, respectively, and σ is the spread of the logon,
Fig. 1. a) Normalized Mean Square Error versus the number
of dictionary elements used in reconstruction, b) Actual and
Reconstructed EEG Signals with K = 10
down to solving for the sparse vector a and then using the
relationship x = Ψa. The sparse vector a can be found by
using the OMP algorithm discussed in [5]. Fig. 2 shows the
average NMSE in reconstruction versus the number of mea-
surements. From this figure, it can be seen that one can recon-
struct the EEG signal with a small amount of error by storing
80 measurements instead of the 750 samples in the original
recording.
0.018

previous work  has shown that any EEG signal can be writ- 0.017
K
ten as x(n) ≈ i=1 ai gi (n) where K is much less than the
Normalized Mean Square Error

number of samples N . 0.016

In this paper, we investigate the sparsity level of EEG sig-

0.015
nals based on the analysis of 750 sample long EEG recordings
of a control subject performing a continuous performance task 0.014
(CPT). The same task is repeated 116 times to produce 116
EEG trials. First, we approximate each EEG recording us- 0.013

ing the elements of the Gabor dictionary over the given time
0.012
range and a frequency range of 1-50 Hz. The normalized
K squared error between the approximated signal, x̂ =
mean 0.011

i=1 ai gi (n) and the original signal is computed and aver-

20 25 30 35 40 45 50 55 60
Number of Measurements per Trial
aged over 116 trials for different values of K. Fig. 1 (a)
shows the average normalized mean squared error (NMSE)
Fig. 2. Normalized Mean Square Error in reconstruction ver-
versus the sparsity level. It can be seen that the NMSE is low
sus the number of measurements
and rolls-off fast as K increases. Fig. 1(b) shows an exam-
ple EEG recording and its reconstruction using 10 elements
of the Gabor dictionary.
3.1. Joint Sparsity of EEG Signals
Using the fact that EEG recordings are sparse, one can use
the CS framework to reconstruct these signals from cK mea- Most of the existing work in EEG signal analysis focuses on
surements. In this paper, the projection matrix, Φ, is chosen the representation of single-trial EEG signals on a given dic-
as i.i.d. Gaussian random vectors. The measurement vector tionary or the representation of the sample average. The re-
y = Φx = ΦΨa, is sparse in the dictionary ΦΨ, since a is sults of the previous section show that individual EEG sig-
a sparse coefficient vector. Therefore, reconstructing x boils nals are sparse in a Gabor frame and can be reconstructed

182
from a small number of projections onto an incoherent ba- 0.019
sis. Since EEG signals are usually recorded over multiple
trials and electrodes, this projection and the following recon- 0.018

struction has to be done on each signal, individually. This 0.017

Normalized Mean Square Error

approach is not practical as it would become computationally 0.016
expensive to construct different projection matrices for each
0.015
signal and to reconstruct them separately. Therefore, it is im-
portant to investigate the existence of a basis or dictionary 0.014

that can efficiently represent a collection of EEG signals since 0.013

the joint sampling and reconstruction of EEG signals would 0.012
prove to be more beneficial in reducing the number of samples
that are collected compared to the individual sampling. 0.011

In this section, we assume the following joint sparsity 0.01

0 5 10 15 20 25 30 35 40
model for a collection of EEG signals, xj , j ∈ {1, 2, . . . , J}, Number of Dictionary Elements

where J = 116 in our case.

Fig. 3. Normalized Mean Square Error in reconstruction ver-
xj = Ψaj , (3) sus the number of dictionary elements used

where each aj is supported on the same subset of the dictio-

nary. Hence, all signals are K-sparse and are reconstructed 0.07
from the same K elements of Ψ, but with arbitrarily different
coefficients. This is a reasonable assumption for EEG record- 0.06
ings since the different sensors measure the same activity with

Normalized Mean Square Error

phase difference and attenuation caused by the signal propa- 0.05
gation in the brain.
In this paper, we consider the 116 trials of EEG activ- 0.04

ity recorded during a continuous performance task. We as-

sume that the different trials, similar to the different elec- 0.03

trodes, measure the same underlying activity with phase re-

settings. We investigate the joint sparsity of 116 trials over 0.02

the Gabor dictionary as illustrated in Fig. 3. From this fig-

ure, it can be seen that 40 elements from the Gabor dictionary 0.01
20 30 40 50 60 70 80
result in an average NMSE that is comparable to the average Number of Measurements

NMSE obtained by choosing 3 dictionary elements per trial

(see Fig. 1 (a)). This means that by using the joint sparsity
we can reduce K from 3 × 116 = 348 to 40 and achieve the
same NMSE.
Since the EEG signals are jointly sparse, we can jointly re-
construct the signals from their individual measurements us-
ing the distributed compressed sensing-simultaneous OMP al-
gorithm discussed in [6]. Fig. 4 illustrates the average NMSE
in the joint reconstruction of the EEG signals versus the num-
ber of measurements. It can be seen that the error rates in Fig.
4 are slightly higher than the ones found for the individual
trial reconstruction in Fig. 2 since the same set of elements
from the Gabor dictionary are used to reconstruct all of the
EEG recordings.
4. ACKNOWLEDGEMENTS
The author thanks Edward Bernat and Scott Sponheim from
the Department of Psychology, University of Minnesota for
sharing their EEG data.
Fig. 4. Normalized Mean Square Error in reconstruction ver-
sus the number of measurements
5. CONCLUSIONS
In this paper, we adapted the compressed sensing framework
for the compression and representation of EEG signals. First,
we showed the sparsity of individual EEG recordings in a Ga-
bor dictionary and used that information to reduce the number
of measurements required to reconstruct the EEG signal. We
then extended this to a joint sparsity framework to further re-
duce the number of measurements. The proposed work can be
extended by modifying the Gabor frame to a chirped Gabor
dictionary to further reduce the sparsity level and the corre-
sponding number of measurements. Moreover, the quality of
the reconstructed signals can be evaluated criteria other than
mean-squared error such as qualitative measures that describe
the quality of diagnosis.

183
 6. REFERENCES

[1] E. Candés and T. Tao, “Near-optimal signal recovery

from random projections: universal encoding strategies,”
IEEE Trans. on Information Theory, vol. 52, pp. 5406–
5425, 2004.

[2] D. Donoho, “Compressed sensing,” IEEE Trans. on In-

formation Theory, vol. 52, pp. 1289–1306, 2006.

[3] M. B. Wakin, J. N. Laska, M. F. Duarte, D. Baron, S. Sar-

votham, D. Takhar, K. F. Kelly, and R. G. Baraniuk, “An
architecture for compressive imaging,” in Proceedings
of IEEE International Conference on Image Processing,
2006, pp. 1273–1276.

[4] M. Lustig, D. Donoho, and J. M. Pauly, “Sparse

MRI: The application of compressed sensing for rapid
MR imaging,” Magnetic Resonance in Medicine, 2007,
preprint.

[5] J. Tropp and A. C. Gilbert, “Signal recovery from par-

tial information via orthogonal matching pursuit,” 2005,
preprint.

[6] M. F. Duarte, S. Sarvotham, D. Braon, M. B. Wakin,

and R. G. Baraniuk, “Distributed compressed sensing of
jointly sparse signals,” in IEEE Proceedings of the 2005
Asilomar Conference on Signals, Systems, and Comput-
ers, 2005, pp. 1537–1541.

[7] M. L. Brown, W. J. Williams, and A. O. Hero, “Non-

orthogonal gabor representation of event-related poten-
tials,” in Proceedings of the 15th Annual International
Conference of IEEE Engineering in Medicine and Biol-
ogy Society, 1993, pp. 314–315.

[8] P. J. Durka and K. J. Blinowska, “A unified time-

frequency parametrization of EEGs,” IEEE Engineer-
ing in Medicine and Biology, vol. 20, no. 5, pp. 47–53,
September/October 2001.

[9] S. Aviyente, E. Bernat, S. Malone, and W. Iacono, “Anal-

ysis of event related potentials using pca and matching
pursuit on the time-frequency plane,” in IEEE Interna-
tional Conference of the Engineering in Medicine and Bi-
ology Society, 2006, vol. 3, pp. 424–427.

184