IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO.

1, JANUARY 2008 87
Performance Evaluation of Neural Network
and Linear Predictors for Near-Lossless
Compression of EEG Signals
N. Sriraam, Member, IEEE, and C. Eswaran, Senior Member, IEEE
AbstractThis paper presents a comparison of the perfor-
mances of neural network and linear predictors for near-lossless
compression of EEG signals. Three neural network predictors,
namely, single-layer perceptron (SLP), multilayer perceptron
(MLP), and Elman network (EN), and two linear predictors,
namely, autoregressive model (AR) and nite-impulse response l-
ter (FIR) are used. For all the predictors, uniform quantization is
applied on the residue signals obtained as the difference between
the original and the predicted values. The maximum allowable re-
construction error is varied to determine the theoretical bound

0
for near-lossless compression and the corresponding bit rate
r
p
. It is shown that among all the predictors, the SLP yields the
best results in achieving the lowest values for
0
and r
p
. The cor-
responding values of the delity parameters, namely, percent of
root-mean-square difference, peak SNR and cross correlation are
also determined. Acompression efciency of 82.8%is achieved us-
ing the SLP with a near-lossless bound
0
= 3, with the diagnostic
quality of the reconstructed EEG signal preserved. Thus, the pro-
posed near-lossless scheme facilitates transmission of real time as
well as ofine EEG signals over network to remote interpretation
center economically with less bandwidth utilization compared to
other known lossless and near-lossless schemes.
Index TermsEEG, near-lossless compression, neural network,
prediction, uniform quantization.
I. INTRODUCTION
T
HE NEED for data storage and transmission through net-
work for telemedicine applications has made a great de-
mand for efcient EEG signal compression techniques. Data
compression, in general, can be lossless, when the signal wave-
form delity is totally preserved or lossy in cases where a
certain amount of distortion or lack of delity in the decom-
pressed data is allowed. For EEG signals, lossless compression
schemes based on linear and neural network predictors have
already been reported [1][8]. Although, many compression
techniques have been reported, the search for new methods con-
tinues, with the aim of achieving greater compression while
preserving the clinical information content in the reconstructed
signal.
Manuscript received July 27, 2006; revised November 30, 2006 and March
30, 2007.
N. Sriraam is with the Department of Information Technology, Sri Sivasub-
ramaniya Nadar (SSN) College of Engineering, 603110 Chennai, India (e-mail:
sriraam@ssnce.ac.in).
C. Eswaran is with the Faculty of Information Technology, Multimedia
University, 63100 Cyberjaya, Malaysia (e-mail: eswaran@mmu.edu.my).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TITB.2007.899497
If we uniformly quantize the error signals such that the mag-
nitude of the reconstructed error lies within a certain bound,
say , improvement in the compression results can be achieved,
without degrading the quality of the signal considerably. Such
a scheme has been referred to as L

-constrained compression
or near-lossless compression, and has been applied for images
[9][11]. For EEGsignals, a context-based near-lossless scheme
with = 1 using linear predictor has been reported [8]. In gen-
eral, a high value of error bound results in degrading the quality
of the reconstructed signal, thereby, losing critical information
important for clinical diagnosis. This can be referred as lossy
compression. It will be useful to identify an appropriate value
of the error bound that will indicate whether a compression
scheme is near-lossless or lossy. In this paper, we have obtained
the theoretical bounds for near-lossless compression of EEGsig-
nals based on the concept reported in [11]. Experiments are con-
ducted using EEGsignals recorded under different physiological
conditions with values ranging from 1 to 10, and the theoreti-
cal near-lossless bound
0
based on the maximum value of error
bound that satisfy the near-lossless criteria reported in [11]
is determined. Three neural network predictors, namely, single-
layer perceptron (SLP), multilayer perceptron (MLP), and El-
man network (EN) [4][7], and two linear predictors, namely,
autoregressive (AR) model [8] and normalized least-mean
square adaptive nite-impulse response (FIR) lter [12], [13] are
considered.
II. NEAR-LOSSLESS SCHEME USING UNIFORM QUANTIZATION
For EEG signal compression, two-stage lossless compression
schemes involving predictor in the rst stage and entropy en-
coder in the second stage have been successfully used [1][8].
The predictor estimates the present value of a sample using
its past samples, and the residues are calculated as the differ-
ence between the original and the predicted samples, which are
generally of a lesser magnitude than the original samples. It is
assumed that both the encoder and the decoder simulate iden-
tical prediction processes [1][8]. The prediction process starts
with the transmission of initial header information consisting of
predictor parameter settings and selected number of input sam-
ple values that is equal to the predictor order. At the receiving
end, the prediction process is repeated and the original input is
recovered by adding the transmitted residues to the predicted
values. The raw EEG input signals are rst divided into blocks
of specic duration. For prediction using neural networks, each
block of samples are trained with neural network using a block-
adaptive training scheme [14] as discussed by the authors [4].
1089-7771/$25.00 2008 IEEE
88 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008
Fig. 1. Near-lossless compression scheme.
The details of transmission procedure are described in [4].
Fig. 1 shows the steps involved in a near-lossless compression
scheme.
If x
k
is the current sample and x
k
is the predicted sample,
then the value of the error (residue) sample is given by
e
k
= x
k
x
k
. (1)
The error sample e
k
is then uniformly quantized [8]
[e
k
]
q
=
_
e
k
+
2 + 1
_
(2)
where . denotes the integer part of the argument and is the
maximum allowable reconstruction error between the original
and the reconstructed signal.
The quantized error signal [e
k
]
q
is then encoded using an
arithmetic encoder [15], [16], and the resultant signal is de-
noted as [e
k
]
qe
. At the receiving end, the reconstructed signal is
obtained
x
k
= x
k
+ [e
k
]
qe
(3)
where
x
k
is the reconstructed signal;
x
k
is the predicted signal at the receiving end;
[e
k
]
qe
is the transmitted error signal.
III. CRITERION FOR NEAR-LOSSLESS COMPRESSION
In this section, the criterion for near-lossless compression of
EEG signals is derived using the approach given in [11]. Let
represent the maximum allowable error between the original
and the reconstructed EEG signal
|x
k
x
k
| (4)
where x
k
and x
k
are original and reconstructed signal values at
time k, respectively.
A zero value of guarantees lossless reconstruction. A
marginal variation of may lead to a near-lossless compression
in which the quality of the reconstructed signal is tightly con-
trolled to avoid any appreciable visual degradation of the EEG
waveform. Let r
L
be the best possible expected bit rate, i.e.,
Shannons rst-order entropy obtained for a lossless EEG com-
pression. Let r
p
represent the practical bit rate achieved using
uniform quantization as shown in (2) for a near-lossless com-
pression. If x
k
is the reconstructed signal obtained as given in
(3), then the difference between the reconstructed and the orig-
inal signal, i.e., x
k
x
k
can take at most 2 + 1 values [11],
and this sequence can be encoded with a expected bit rate r
e
,
which is given by [11]
r
e
log
2
(2 + 1) (5)
and hence, we have [11]
r
L
r
p
+ r
e
. (6)
Combining (5) and (6), we get
r
p
r
L
log
2
(2 + 1). (7)
It has been shown in [11] that (7) can be generalized in such a
way that for a small value of , log
2
(2 + 1) can be replaced by
the entropy of the reconstruction error H
e
r
P
r
L
H
e
. (8)
The term on the right-hand side of (8), i.e., r
L
H
e
represents
the theoretical bound for the bit rate that can be achieved for a
near-lossless compression with uniform quantization, and it is
denoted as r
NLC
. The maximum value of , which satises (8) is
dened as the theoretical bound for near-lossless compression
of , and is denoted as
0
.
IV. PERFORMANCE EVALUATION
For performance evaluation, EEG signals recorded under
three different physiological conditions are used. Dataset1
(DS1) consists of EEG signal recorded under epileptic con-
dition with sudden seizures with a sampling rate of 256 Hz with
16 bit accuracy obtained from hospital Universiti Kebangsaan
Malaysia (UKM) [7], [17]. Dataset2 (DS2) consists of normal
EEG signal with eyes open and closed recorded using BIOPAC
data acquisition system with a sampling rate of 250 Hz with 16
bit accuracy [7], [17], [18]. DS2 has been recorded in Multime-
dia University biomedical laboratory. Dataset3 (DS3) contains
EEGsignals recorded during the occurrence of epileptic seizures
exhibiting ictal activity with a sampling rate of 173.61 Hz
with 12 bit accuracy [19]. For the experimental purpose, 6 min
(360 s) recordings of DS1, DS2, and DS3 are considered, and
are divided into four blocks with each block consisting of 90 s
duration of samples that are used for further prediction. Fig. 2
shows the samples of DS1, DS2, and DS3 with 90 s duration of
recordings.
For prediction using neural networks, three models, namely,
SLP, MLP, and EN with their optimal congurations as reported
SRIRAAM AND ESWARAN: PERFORMANCE EVALUATION OF NEURAL NETWORK AND LINEAR PREDICTORS 89
Fig. 2. Sample recordings of EEG under different physiological conditions.
TABLE I
OPTIMAL CONFIGURATION OF NEURAL NETWORKS
[5] are used and the details are shown in Table I [20]. For linear
prediction, a fth-order AR model [7], [17] and a fth-order
FIR [12], [13] are used. In order to estimate the AR parameters,
LevinsonDurbins method as reported in [8] is used.
The performances of the different predictors for the near-
lossless compression of EEG signals are evaluated using the
parameters bit rate [bits per sample (BPS)], percent of root-
mean-square difference (PRD), peak SNR (PSNR), and cross
correlation (CC).
PRD is dened as [21], [22]
PRD =

N
k=1
(x
k
x
k
)
2

N
k=1
(x
k
)
2
100 (9)
where x
k
and x
k
are original and reconstructed signals and N is
the length of the window over which the PRD is calculated.
PSNR is dened as [23], [24]
PSNR = 20 log
_
max(x
k
)
RMSE
_
(10)
where max(x
k
) is the maximum value of the original EEG
signal and RMSEis the root-mean-square error, which is dened
as
RMSE =

_
1
N
(PRD)
2
N

k=1
(x
k
)
2
. (11)
TABLE II
NEAR-LOSSLESS COMPRESSION RESULTS OBTAINED FOR DIFFERENT
PREDICTORS
CC denotes the statistical correlation between two sig-
nals [25]. The correlation between the original and the re-
constructed signal is measured by CC, which is dened
as
CC =
cov(x
k
x
k
)
_
var (x
k
)var( x
k
)
(12)
where cov(x
k
x
k
) is the covariance between x
k
and x
k
,
and var(x
k
) and var( x
k
) are the variances of x
k
and x
k
,
respectively.
Table II shows the compression results obtained using
neural network and linear predictors for the experimental
datasets at different values of . For the datasets DS1, DS2,
and DS3, the compression results for each block are deter-
mined rst, and the average values obtained using all the
datasets corresponding to all four blocks are reported in
Table II.
From Table II, it is seen that as increases, the practical
bit rate r
P
for near-lossless compression decreases. The values
of the theoretical
0
, which satisfy (8) for different predictors
obtained from Table II are as follows:

0
= 3 (for SLP and AR)

0
= 4 (for MLP and FIR)

0
= 5 (for EN).
Among the different predictors, SLP yields the best results
with lowest values of
0
and r
P
. The original and the recon-
structed signals obtained using SLP at different values of for
DS1 with a specic block of 90 s duration are shown in Fig. 3(a)
(e). Only traces of samples corresponding to the duration
6090 s are shown.
From Fig. 3(a)(c), it is seen that the increase in results in
degrading the quality of the reconstructed EEGsignal which can
be assessed qualitatively through visual inspection. Fig. 4(a)(c)
shows the average values of PRD, PSNR, and CC, respectively,
at different values of for different predictors.
It can be seen fromFig. 4 that as increases, the quality of the
reconstructed signal decreases and SLP yields the best delity
90 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008
Fig. 3. Original and reconstructed signal using SLP with DS1. (a) = 1. (b) = 3. (c) = 5. (d) = 7. (e) = 9.
measures for all values of . Table III shows the values of r
p
and the values of the delity parameters corresponding to
0
for
different predictors.
From Table III, it is clear that among the different predictors,
SLP yields the best results with respect to the bit rate and the
delity measures.
V. DISCUSSION
The reliability of the near-lossless scheme using neural net-
work and linear predictors has been assessed based on the recon-
structed EEG signals qualitatively and quantitatively. It can be
seen from Fig. 3 that the visual inspection of the reconstructed
EEGsignals reveals that the near-lossless bound
0
= 3 obtained
SRIRAAM AND ESWARAN: PERFORMANCE EVALUATION OF NEURAL NETWORK AND LINEAR PREDICTORS 91
Fig. 4. (a) Variation of PRDwith . (b) Variation of PSNRwith . (c) Variation
of CC with .
TABLE III
BOUNDARY VALUES FOR NEAR-LOSSLESS COMPRESSION
using SLP preserves the diagnostic accuracy required for clin-
ical diagnosis. The clinical validity of the resulting signal can
also be assessed from the PRD, PSNR, and CC values shown in
Table III and Fig. 4. For SLP, with
0
= 3, these values are 1.25,
39.26, and 0.92, respectively. For situations that demand still
better quality of the reconstructed signal, one can apply the pro-
posed SLP-based method with < 3. Similarly for situations
where the requirement on the quality of the reconstructed signal
can be relaxed, one can use values of > 3. Thus, the proposed
method allows the user to enforce a tight control on the quality
of the reconstructed signal. Although SLP and AR obtain the
same near-lossless bound
0
= 3, the quantitative analysis of the
reconstructed EEG signal using delity parameters, as shown
in Fig. 4 and Table III, reveals that SLP yields better delity
results. It is well known that the two factors, namely, BPS and
PRDdecide the quality on demand specications, namely, band-
width constraints and reconstructed signal quality, respectively,
for the transmission of EEG signals. Fig. 5 shows the variation
of PRD with respect to the practical bit rate r
p
dened in BPS
for different predictors by varying the value from 1 to 10.
It is seen from Fig. 5 that SLP requires less BPS compared
to AR for all the values of . The compression performance
of the proposed near-lossless scheme is compared with other
known schemes reported in the literature [1][8]. It may be
mentioned that as the EEG data considered for analysis in the
various schemes are different, an exact comparison cannot be
performed. Table IV shows the compression efciency [CE (in
percent)], as dened in (13) [1], obtained using the lossless and
near-lossless schemes reported in [1][8] as well as the near-
lossless scheme discussed in this paper, which is given by
CE = 100
L
Orig
L
Comp
L
Orig
(13)
where L
Orig
is the original le length (in bits) and L
Comp
is the
compressed le length (in bits).
From Table IV, it can be concluded that the proposed near-
lossless scheme using SLP is superior to other schemes with
regard to the CE. A saving of 3.16 bits/sample is achieved using
the proposed near-lossless scheme using SLP when compared
to the lossless scheme using SLP reported by authors in [7].
92 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008
Fig. 5. Variation of PRD with BPS.
TABLE IV
COMPARISON OF CE OF PROPOSED NEAR-LOSSLESS SCHEME WITH EXISTING
SCHEMES
VI. CONCLUSION
A performance evaluation of neural network and linear pre-
dictors for near-lossless compression of EEG signals has been
carried out in this paper. Uniform quantization has been applied
to the residue signals, and the maximumallowable error between
the original and the reconstructed signal has been varied to de-
termine the theoretical bound
0
for the near-lossless compres-
sion of EEG signals. Three neural network predictors, namely,
SLP, MLP, and EN, and two linear predictors, namely, ARmodel
and FIR lter have been used. Experiments were performed us-
ing EEG signals recorded under three different physiological
conditions. It has been shown that among all the predictors, the
SLP yields the best results for the near-lossless compression of
EEG signals. Due to the nonlinear nature of the neural network,
the variance of the prediction error of SLP was found to be
less compared to that of the AR linear predictor, which results
in increased predictive gain. The generalization, self-learning,
and robustness properties of neural networks makes it suitable
candidate for nonlinear signal prediction problems than linear
predictors. Although SLP and AR attains the same near-lossless
bound
0
= 3, the qualitative and quantitative evaluation of the
reconstructed EEG signals reveals that the SLP maintains the
diagnostic quality of the reconstructed EEG signal with low
PRD and high PSNR values (1.25 and 39.26 dB).
With = 3, the proposed SLP scheme yields a higher CE of
82.8% compared to other known compression schemes. A sav-
ing of 3.16 bits/sample indicates that the proposed near-lossless
scheme is cost effective for EEG database management as well
as for telemedicine applications as the demand on memory space
and network bandwidth is reduced considerably. The quality
or the clinical validity of the reconstructed signal is assessed
based on the visual analysis and delity parameters. In the pro-
posed scheme, by varying the value of , the user can enforce
a tight control on the quality of the reconstructed signal, from
lossless (exact reconstruction) to lossy. This exibility allows
the user to choose an appropriate value of (above or below
the value of the near-lossless theoretical bound
0
) depending
on the clinical requirement of the quality of the reconstructed
signal.
ACKNOWLEDGMENT
The authors would like to thank the reviewers for their valu-
able comments.
REFERENCES
[1] G. Antoniol and P. Tonella, EEG data compression techniques, IEEE
Trans. Biomed. Eng., vol. 44, no. 2, pp. 105114, Feb. 1997.
[2] N. Magotra, G. Mandyam, M. Sun, and J. W. McCoy, Lossless compres-
sion of electroencephalographic data, in Proc. IEEE Int. Circuits Syst.
Symp., 1996, vol. 2, pp. 313315.
[3] F. Bartolini, V. Cappellini, S. Nerozzi, and A. Mecocci, Recurrent neu-
ral network predictors for EEG signal compression, in Proc. Int. Conf.
Acoustic Speech Signal Process., 1995, vol. 5, pp. 33953398.
[4] N. Sriraam, R. Kannan, and C. Eswaran, Lossless compression of EEG
data using neural network predictors, in Proc. 9th Int. Conf. Neural Inf.
Process., 2002, vol. 5, pp. 20462048.
[5] N. Sriraam and C. Eswaran, EEG signal compression using optimally
congured neural network predictors, in Proc. IEEE Int. Conf. Med.
Signal Inf. Process., 2004, pp. 378382.
[6] N. Sriraamand C. Eswaran, Performance evaluation of two-stage lossless
compression of EEGsignals, Int. J. Signal Process., vol. 1, no. 2, pp. 94
97, Dec. 2004.
[7] N. Sriraam and C. Eswaran, Lossless compression algorithms for EEG
Signals: A quantitative evaluation, in Proc. IEEE/EMBS 5th Int. Work-
shop Biosignal Interpretation, 2005, pp. 125130.
[8] N. Memon, X. Kong, and J. Cinkler, Context-based lossless and near-
lossless compression of EEGsignals, IEEE Trans. Inf. Technol. Biomed.,
vol. 3, no. 3, pp. 231238, Sep. 1999.
[9] X. Wu and P. Bao, L

constrained high-delity image compression via

adaptive context modeling, IEEE Trans. Image Process., vol. 9, no. 4,
pp. 536542, Apr. 2000.
[10] P. Bao and X. Wu, L

constrained near-lossless image compression

using weighted nite automata encoding, Comput. Graph., vol. 22,
no. 2, pp. 217223, 1998.
SRIRAAM AND ESWARAN: PERFORMANCE EVALUATION OF NEURAL NETWORK AND LINEAR PREDICTORS 93
[11] M. Klimesh, Quantization considerations for distortion-controlled
data 324-compression, Jet Propulsion Laboratory (JPL), Pasadena
CA, Telecommun. Mission Oper. Progress Rep. 42-139, Nov. 1999,
pp. 138.
[12] S. Haykin, Adaptive Filter Theory, 4th ed. Englewood Cliffs, NJ:
Prentice-Hall, 2001.
[13] Filter Design Toolbox 3.0 (For Use With Matlab), The Math Works, Natick
MA, 2000.
[14] R. Logeswaran and C. Eswaran, Neural network based lossless coding
schemes for telemetry data, in Proc. IEEE Int. Conf. Geosci. Remote
Sens., 1999, vol. 4, pp. 20572059.
[15] J. W. McCoy, N. Magotra, and S. Stearns, Lossless predictive coding,
in Proc. IEEE Midwest Symp. Circuits Syst., Los Angeles, CA, 1995,
pp. 927930.
[16] S. Stearns and D. Samuel, Arithmetic coding in lossless waveform com-
pression, IEEE Trans. Signal Process., vol. 43, no. 8, pp. 18741879,
Aug. 1995.
[17] N. Sriraam and C. Eswaran, Lossy compression of EEG Signals using
linear and neural network predictors, presented at the Int. Conf. Comput.
Intell., Robot. Auton. Syst., Singapore, Dec. 2005.
[18] Bipoac Owners Guide. (2001). [Online]. Available: www.biopac.com.
[19] R. G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, and C.
E. Elger, Indications of nonlinear deterministic and nite-dimensional
structures in time series of brain electrical activity: Dependence on record-
ing region and brain state, Phys. Rev. E, vol. 64, pp. 061907-1061907-8,
2001.
[20] H. Demuth and M. Beale, Neural Network Toolbox (For Use With Matlab).
Natick, MA: The Math Works, 2000.
[21] M. A. Sabah, A. Al-Shrouf, and M. Abo-Zahhad, ECG data compression
using optimum non-orthogonal wavelet transform, Med. Eng. Phys.,
vol. 22, no. 1, pp. 3946, 2000.
[22] J. L. Cardenas-Barrera and J. V. Lorenzo, Mean-shape vector quantizer
for ECG signal compression, IEEE Trans. Biomed. Eng., vol. 46, no. 1,
pp. 6270, Jan. 1999.
[23] A. N. Netravali and B. G. Haskell, Digital Pictures: Representa-
tion, Compression and Standards, 2nd ed. New York: Plenum Press,
1995.
[24] V. Bhaskaran and K. Konstantinides, Image and Video Compression
Standards, Algorithms and Architectures. New York: Springer-Verlag,
1997.
[25] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete Time Signal
Processing, 2nd ed. Englewood Cliffs, NJ: Prentice-Hall, 1999.
N. Sriraam (M01) received the B.E. dgeree in
electronics and communication engineering (ECE)
from the National Engineering College, Tamil Nadu,
India, in 1996, the M.Tech. degree (with distinc-
tion) in biomedical engineering from Manipal In-
stitute of Technology (MIT), Manipal, India, in
2000, and the Ph.D. degree in information tech-
nology from the Multimedia University, Cyberjaya,
Malaysia, in the area of biomedical signal processing,
in 2007.
He is currently an Assistant Professor in the De-
partment of Information Technology, Sri Sivasubramaniya Nadar (SSN) College
of Engineering, Chennai, India. He is the author or coauthor of 20 articles pub-
lished in journals and has been involved in several sponsored research projects.
His current research interests include biomedical signal and image processing,
data mining, neural networks.
Dr. Sriraam is a member of the IEEE Engineering in Medicine and Biology
(EMB) Society and has secured the University First Rank and was awarded the
gold medal in M.Tech.
C. Eswaran (SM90) received the B.Tech., M.Tech.,
and Ph.D. degrees from the Indian Institute of Tech-
nology, Madras, India, in 1964, 1967, and 1974, re-
spectively.
He is currently a Professor in the Faculty of Infor-
mation Technology, Multimedia University, Cyber-
jaya, Malaysia, where he is also the Chairman of the
Center of Excellence, Multimedia Computing. He
has also been a Professor in the Department of Elec-
trical Engineering, Indian Institute of Technology,
Madras. He was a Research Fellow/Visiting Faculty
in Ruhr University, Bochum, Germany; Concordia University, Montreal, QC,
Canada; University of Victoria, BC, Canada; and Nanyang Technological Uni-
versity, Singapore. He is the author or coauthor of more than 120 research
papers published in reputed international journals and conferences. He has also
been involved in several sponsored research projects and served as an Industrial
Consultant.