IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO.
3, MARCH 2000
Nonlinear Adaptive Filtering of Stimulus Artifact
Richard Grieve, Philip A. Parker*, Senior Member, IEEE, Bernard Hudgins, Member, IEEE, and
Kevin Englehart, Member, IEEE
Abstract—Noninvasive measurements of somatosensory evoked
potentials have both clinical and research applications. The elec-
trical artifact which results from the stimulus is an interference
which can distort the evoked signal, and introduce errors in
response onset timing estimation. Given that this interference is
synchronous with the evoked signal, it cannot be reduced by the
conventional technique of ensemble averaging. The technique of
adaptive noise cancelling has potential in this regard however, and
has been used effectively in other similar problems. An adaptive
noise cancelling filter which uses a neural network as the adaptive
element is investigated in this application. The filter is imple-
mented and performance determined in the cancelling of artifact
for in vivo measurements on the median nerve. A technique of
segmented neural network training is proposed in which the
network is trained on that segment of the record time window
which does not contain the evoked signal. The neural network is
found to generalize well from this training to include the segment
of the window containing the evoked signal. Both quantitative
and qualitative measures show that significant stimulus artifact
reduction is achieved.
Index Terms—Adaptive, artifact, cancelling, evoked, filter, re-
sponse, stimulus.
I. INTRODUCTION
S OMATOSENSORY evoked potentials (SEP’s) are used
widely in clinical and research applications, and include
peripheral, spinal, and cortical potentials. For reasons of
convenience, comfort, and safety, the preferred measurement
techniques are noninvasive using surface electrical stimulation
and recording electrodes. The recorded SEP is thus small in
amplitude—typically on the order of a microvolt peak [1].
While an electrical stimulus evokes the desired response from
the nervous process of interest, it also results in an interfering
potential referred to as the stimulus artifact (SA).
Due to a number of factors, the stimulus artifact is a partic-
ularly troublesome form of interference. First, the amplitude of
the SA is typically orders of magnitude larger than the SEP, [2].
Second, the SA is synchronous or coherent with the SEP and
thus cannot be reduced by ensemble averaging—the common
method for reducing other forms of interference, [3]. Third, in
most cases the SA overlaps the SEP in both the time and fre-
quency domains, such that conventional time windowing and
frequency filtering are not capable of removing the SA without
Manuscript received September 11, 1998; revised October 8, 1999. This work
was supported in part by the NSERC under a Postgraduate Scholarship and
Grant OGP0004445. Asterisk indicates corresponding author.
R. Grieve B. Hudgins, and K. Englehart are with the Department of Electrical
and Computer Engineering and Institute of Biomedical Engineering, University
of New Brunswick, Fredericton, NB, E3B5A3 Canada.
*P. A. Parker is with the Department of Electrical and Computer Engineering
and Institute of Biomedical Engineering, University of New Brunswick, Fred-
ericton, NB, E3B5A3, Canada (e-mail: pap@unb.ca).
Publisher Item Identifier S 0018-9294(00)01772-9.
0018–9249/00$10.00 © 2000 IEEE
389
distortion of the SEP, [4]. Fourth, the mechanisms by which
the electrical stimulus is coupled into the recording are varied,
complex and distributed, making minimization of the SA at the
source problematic [5].
Given that the conventional methods are not effective for SA
reduction in SEP measurements, alternative approaches have to
be considered. A technique which has potential in this regard
is adaptive noise cancelling (ANC). In this approach an esti-
mate of the SA is obtained via an adaptive filter and the esti-
mate subtracted from the composite SA plus SEP signal. ANC
filters have been used in other biological signal acquisition ap-
plications. Ye and Choy [6] used such a filter for the reduction of
respiratory artifact in rheopneumography measurements. Sada-
sivan and Narayana [7] developed an ANC filter for the reduc-
tion of electrooculogram in electroencephalography measure-
ments. Thakor and Zhu [8] have used an ANC filter for the re-
moval of the QRS complex in ECG in order to improve detec-
tion of the -wave. Suzuki et al. [9] developed a real-time ANC
filter for suppression of ambient noise in lung sound measure-
ments.
In the application of noise cancelling to SA reduction,
McGill et al. [4] estimated the artifact waveform from sub-
threshold stimulation and subtracted the estimate from the
suprathreshold stimulus response. Nonlinear adaptive ANC
filters were not considered in this study. Work by one of the
present authors investigated the application and performance
of linear and nonlinear finite impulse response (FIR) ANC
filters for SA reduction in SEP measurements [10], [11]. The
performance of linear FIR ANC filters was found to be rather
poor [10] suggesting that the input–output relationship of tissue
to electrical stimulus is nonlinear. Indeed the source of this
nonlinearity is, at least in part, the nonlinear voltage–current
characteristic of the skin [12]. A second-order truncated
Volterra series nonlinear FIR ANC filter was shown to have
substantially improved performance [10].
While this filter achieves good SA cancellation, it has two
serious limitations: 1) it cannot model nonlinearities of order
higher than two and 2) the number of filter coefficients grows
as the square of the FIR filter length. An alternative ANC filter
that does not suffer from these limitations is the neural network
(NN) configuration. The objectives of this paper are to investi-
gate the application and performance of the NN ANC filter for
the reduction of SA in SEP measurement.
II. NN ANC FILTER
The NN configuration has been used in ANC filters for other
applications requiring cancellation of different types of interfer-
ence. Yuwaraj et al. [13] applied the NN ANC filter for cancel-
lation of interference in otoacoustic emission measurements. In
390 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 3, MARCH 2000
Fig. 1. Adaptive noise cancelling (AN) filter structure.
Fig. 2. Feedforward neural network structure.
the first application to SA reduction, Grieve et al. [14] inves-
tigated the pi–sigma NN configuration with some success but
concluded that a feedforward NN configuration would be more
appropriate. The feedforward multilayer perceptron (MLP) NN
will be considered in this paper for the ANC filter and SA can-
cellation.
The general configuration for the ANC filter is given in Fig. 1.
The primary channel input is the addition of evoked re-
sponse , artifact , and uncorrelated noise . The reference
channel input is the addition of artifact , correlated with
through the system , and uncorrelated noise . The adap-
tive filter estimates from and is adapted by the error
through an appropriate training algorithm. Following conver-
gence of , the ANC filter output is the minimum mean
square error estimate of .
Different structures can be used for the adaptive filter,
, and perhaps the most common is the linear FIR filter.
However the structure of interest here is the NN adaptive
filter because of its ability to generalize from training to test
data, and to model nonlinear transfer functions of arbitrary
order. The general structure of a feedforward MLP neural
network is given in Fig. 2. For the implementation in this
paper, samples, , from form the
-dimensional input feature vector for the input layer. The
input layer is fully connected to the hidden layer through the
weights , where is the number of
hidden nodes. The hidden layer is connected to the output layer
through the weights , and hyperbolic
tangents are used for activation functions. The bias inputs and
associated weights, , are included in practice in order to
accommodate dc offsets.
During the training phase, the NN weights are updated, from
small randomly selected initial values, using the back-propaga-
tion algorithm [15]. The update equations are given by
and
where
is the time index;
is the learning rates for layer 1;
is the learning rate for layer 2;
is the derivative of the activation function evaluated at
;
is the cancellation error at time index .
Appropriate values for , and are determined by per-
formance and stability criteria to be considered in Section IV as
part of two in vivo experiments.
These experiments require the acquisition of SEP and SA
data, and Section III describes the acquisition methodology.
III. METHODOLOGY
A block diagram of the measurement system for acquiring
SEP from the median nerve is shown in Fig. 3. Median nerve
sensory fibers are stimulated at the finger and SEP acquired at
the wrist using surface Ag/AgCl electrodes. The resulting SEP
and SA are amplified over the bandwidth 10 Hz to 10 kHz,
sampled at a rate of 25.6 kHz, and the digital data stored for
subsequent off-line processing. Following each stimulus, a data
window of 10 ms or 256 samples is recorded. The Bruel and
Kjaer 2032 signal analyzer is used in an on-line ensemble aver-
aging mode in order to confirm the presence of SEP during the
recording session. In order to ensure no SEP crosstalk between
the primary and reference inputs to the ANC filter, a necessary
condition for an ANC filter, a double-stimulus single-channel
technique is used. In this technique a series of alternating sub
and supra threshold stimuli are applied to the finger. The re-
sponses to the subthreshold stimuli are used as the reference
data, and those to the suprathreshold as the primary data. Thus
no channel crosstalk is possible. This technique also has the ad-
vantage of requiring only a single data acquisition channel and
associated hardware.
Two Grass S11 stimulators with isolators are used to generate
the series of stimuli. Fig. 4 shows a typical series of stimuli and
acquired data for this technique.
IV. EXPERIMENT 1
In order to optimize the NN parameters and evaluate the NN
ANC performance, the first experiment acquired data without
the presence of SEP. In this way the cancellation performance
can be examined over the full SA duration, including the time
interval in which the SA would otherwise overlap the SEP. If the
SEP were present and unknown it would be impossible to pre-
cisely evaluate the filter’s performance in cancelling the SA. At
GRIEVE et al.: NONLINEAR ADAPTIVE FILTERING OF STIMULUS ARTIFACT
Fig. 3. Stimulation and acquisition system.
Fig. 4. (a) Alternating sub and supra electrical stimuli and (b) resulting SA and SEP.
the same time the network parameters can be studied and op-
timized—in particular network size and training data window
will be investigated. The methodology of Section III was used to
collect 300 primary and reference records. However, for this ex-
periment both reference and primary stimuli were subthreshold,
thus ensuring no SEP in either channel.
Because the SA and SEP are correlated, it is essential to
ensure that during NN training with SEP present, the primary
channel training data not contain SEP. Otherwise the NN filter
will adapt to cancel not only SA but also the SEP. Thus for
391
training, the record time windows were divided into two seg-
ments, and , such that segment contains no SEP and ,
which is the remainder, contains the SEP. This training method
is referred to as segmented-training. Fig. 5 shows an example of
the segments and . Due to the finite conduction velocity of
the nerve fiber, there is always a time delay between the stimulus
application and the onset of the SEP. This interval determines
the magnitude of the time segment . The value of for prac-
tical application can thus be based on the minimum expected
nerve conduction delay between stimulus and SEP arrival at the
392
Fig. 5. The P and P segments for segmented training. The filter is trained on
the P segment where no SEP is present.
recording electrodes. For the median nerve and this measure-
ment setup, a value of 40 samples or 1.56 ms is appropriate. In
other applications an appropriate value can always be esti-
mated. It is important that this estimation and segmentation be
carefully done as the algorithm training is sensitive to the par-
titioning—if the segment is too short poor generalization of
the filter will result and if too long some cancellation of the SEP
can occur.
The performance measures for the AN filter relate to residual
SA in the filter output as a ratio to the SA in the primary channel.
Specifically the measures are , and as defined by
and
where
is the SA component at time-index in the pri-
mary;
is the SA component at time-index in the output;
is the standard deviation of the random variable ;
is a time segment of the record window;
is a time segment containing the remainder of the
record window;
is a measure of the reduction of the SA peak;
is the measure of the power reduction of the SA in
segment ;
is the measure of the power reduction of the SA in
segment .
Since the NN was trained on the segment , which contains
only a portion of the SA, the ability of the NN to generalize and
accurately generate from in both and is critical to
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 3, MARCH 2000
the AN filter performance. Thus the performance measure is
of particular significance.
The 300 acquired response records were divided randomly
into a 150-record training set and a 150-record test set. The pri-
mary and reference inputs used to train the NN were generated
by averaging the training set data. These inputs were made up
of the first 40 data points, preceded by a 20 point zero pad.
The training was continued until the incremental change in the
mean squared error with each iteration was not significant, indi-
cating convergence of the network. The NN size was optimized
by varying the number of input and hidden nodes, training the
network for each combination, and determining the values of
, and for the test set. The number of input and hidden
nodes were varied over 4–20, and 2–14, respectively. Data were
collected and processed for six subjects, and the results for one,
which is typical of the 6, are shown in Fig. 6. After considering
the results for all subjects, and keeping in mind that is of
particular importance here, the combination of 12 input nodes
and six hidden nodes was selected as optimum for further im-
plementations. The low value is due to the fact that the SA
power in of the primary channel is comparable to the uncor-
related channel noise power, hence limiting the maximum value
that can attain.
To determine the generalization ability of the filter, the length
of the training section was varied over 20–100 data points from
a record length of 256 points, the NN trained, and the perfor-
mance measure calculated . The results are shown
in Fig. 7. Again is the measure of most significance since it
measures the performance in , and hence the ability of the NN
to generalize. The value of is fairly constant over the training
range tested indicating that the NN is generalizing well (this ob-
servation will be considered in more detail in Section VI). Also
shown in Fig. 7 is the performance of a nonlinear FIR AN filter
with 10 taps [10] trained with the recursive least squares algo-
rithm, (NRLS), and operating on the same data. For this filter,
generally continues to increase with indicating that its gen-
eralizing ability is not as strong as the NN for in the range
necessary for practical application, i.e., in this case.
V. EXPERIMENT 2
In this experiment the NN ANC filter of Section IV was used
to cancel SA in SA plus SEP signal from six subjects. The
methodology of Section III, using both sub and supra threshold
stimuli as per Fig. 4, was used to acquire the data. As in Sec-
tion IV, 300 primary and reference records were acquired, and
these were randomly split into a 150-record training set and a
150-record test set. In order to test the practical effectiveness of
the filters, the final outputs were obtained by filtering each of
the 150 records in the test set, and averaging the 150 results.
The primary and reference inputs used to train the NN were
generated by averaging the training set data. These inputs were
made up of the first 40 data points, preceded by a 20 point zero
pad. The training was continued until the incremental change
in the mean squared error with each iteration was not signifi-
cant, indicating convergence of the network. The resulting filter
was then used to filter the 150 records in the test set, and the
150-record filter output was averaged to obtain the final output.
GRIEVE et al.: NONLINEAR ADAPTIVE FILTERING OF STIMULUS ARTIFACT
Fig. 6. Network size test on subthreshold stimulus data showing performance as a function of the number of input and hidden nodes.
Fig. 7. Performance  versus length of training segment.
The training of the NRLS filter was slightly different. It was
found that when the NRLS filter is trained on averaged data, the
kernel values upon which it converges are large numbers. If a
signal is filtered that is slightly different than that used to train
393
the filter, the output contains a very high residual interference.
To prevent the NRLS filter from converging on large kernel
values, it was trained on raw data, rather than averaged data. All
150 primary and reference pairs in the training set were used
to train the filter. Each record was composed of a 20 point zero
pad followed by 40 points of data. The resulting filter was then
used to filter the 150 records in the test set, and the 150-record
filter output was averaged to obtain the final output. It should
be noted that the performance measure cannot be used in this
experiment. Since the SEP is present and unknown, the residual
SA following cancellation cannot be evaluated quantitatively.
Therefore performance evaluation in the segment must be
a qualitative observation of the SA reduction in the overlap re-
gion, relative to the SEP.
Table I gives and values for the two adaptive filters,
NN and NRLS FIR, used in the AN filter over the six subjects.
It is evident that both give substantial reductions in the SA over
the segment . It is also evident that the NN gives a somewhat
larger reduction which is to be expected given that the NRLS is
able to model second order nonlinearities only while the NN can
394 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 47, NO. 3, MARCH 2000

TABLE I 35 dB averaged over six subjects. Since the NN can model a

PERFORMANCE MEASURES  AND 
EXPERIMENT 2 OVER SIX SUBJECTS
Fig. 8. SA cancellation results from Experiment 2 for subjects a; b, and c.
FOR
nonlinearity of arbitrary order it was able to achieve a larger
reduction than an NRLS FIR AN filter based on the truncated
second order Volterra model. It was also demonstrated that the
NN is able to provide better generalization from the to the
segments. The SA cancellation in the interval was shown
qualitatively to be excellent.
Fig. 7 demonstrates the power of the NN in generalizing, and
identifying/modeling the nonlinear system between the primary
and reference channels. The fact that is relatively constant
over the length of investigated indicates that it is able to cap-
ture and model the nonlinear system well with as few as 20
samples per evoked response. For this reason, the NN AN filter
trained in the segment is able to provide excellent SA can-
cellation in the segment. In contrast, the NRLS AN filter
requires a of approximately 70 samples to achieve the same
performance, by which time the SA is typically overlapping the
SEP. The difference is no doubt due to the second order nonlin-
earity modeling limitation of the NRLS, as well as the structural
differences in the NN and FIR adaptive filters. The dip in seen
for both filters as increases is likely due to the fact that in the
early part of the SA is very large compared to the uncorre-
lated noise, and hence the adaption converges well. However
in the later part of the SA falls rapidly to values comparable
to the uncorrelated noise and the adaption and convergence are
disturbed. A corresponding drop in performance occurs until
is sufficiently large to allow the adaption to recover.
From the results of Table I and the small residual SA of Fig. 8,
it is clear that the NN adaptive filter has converged well and
has identified/modeled the nonlinear system well. When
considering training and convergence times, it is noted that in
the case where the SA is stationary, the NN can be trained, the
coefficients fixed, and SEP measurements made on-line or off-
line. In the case of nonstationary SA, the NN adaptive filter can
be allowed to track SA changes over time by using appropriate
values for and in the update algorithm, [11]. Tracking
and convergence properties of the filter in this mode need to be
investigated further in a continuation of this work.

model nonlinearities of arbitrary order. A paired -test shows a
significant difference in their performances at the 5% signifi-
cance level.
Fig. 8 shows the results of the SA cancellation for the NN and
NRLS FIR adaptive filters for three subjects which are typical of
the six subjects. These figures allow a qualitative assessment of
the cancellation in the segment where the SA and
SEP overlap, as well as in the segment . The relative
abilities of the NN and NRLS adaptive filters to generalize from
to are particularly evident in Fig. 8(b).
VI. CONCLUSIONS AND DISCUSSION
The overlap of SA with SEP can cause significant signal dis-
tortion and onset timing errors. An ANC filter with a NN as
the adaptive element was shown to provide substantial SA re-
duction. An MLP with 12 input and six hidden nodes, and seg-
mented training achieved a SA reduction in the segment of
REFERENCES
[1] C. D. McGillem, J. I. Aunon, and D. G. Childers, “Signal processing in
evoked potential research: Applications of filtering and pattern recogni-
tion,” CRC Crit. Rev. Bioeng., vol. 9, pp. 225–265, Oct. 1981.
[2] S. B. Harrison and D. F. Lovely, “Identification of noise sources in sur-
face recordings of spinal somatosensory evoked potentials,” Med. Biol.
Eng. Comput., vol. 33, pp. 299–305, 1995.
[3] D. Regan, Human Brain Electrophysiology. New York: Elsevier Sci-
ence, 1989.
[4] K. C. McGill, K. L. Cummins, B. B. Dorfman, L. J. Berlizot, K. Luetke-
meyer, D. G. Nishimura, and B. Widrow, “ On the nature and elimination
of stimulus artifact in nerve signals evoked and recorded using surface
electrodes,” IEEE Trans. Biomed. Eng., vol. BME-29, pp. 129–137, Feb.
1982.
[5] R. N. Scott, L. McLean, and P. A. Parker, “Stimulus artifact in so-
matosensory evoked potential measurement,” Med. Biol. Eng. Comput.,
vol. 35, pp. 1–5, 1997.
[6] J. Ye and T. T. Choy, “On-line respiratory artifact removal via adap-
tive FIR filters in rheopneumographic measurements,” Med. Biol. Eng.
Comput., vol. 32, pp. 620–624, 1994.
[7] P. K. Sadasivan and D. N. Dutt, “A nonlinear estimation model for adap-
tive minimization of EOG artefacts from EEG signals,” Int. J. Biomed.
Comp., vol. 36, pp. 199– 207, 1994.
GRIEVE et al.: NONLINEAR ADAPTIVE FILTERING OF STIMULUS ARTIFACT
[8] N. V. Thakor and Y. S. Shu, “ Applications of adaptive filtering to ECG
analysis: Noise cancellation and arrhythmia detection,” IEEE Trans
Biomed. Eng., vol. 38, pp. 785–793, Aug. 1991.
[9] A. Suzuki, C. Sumi, K. Nakayama, and M. Mori, “Real-time adaptive
cancelling of ambient noise in lung sound measurements,” Med. Biol.
Eng. Comput., vol. 33, pp. 704–708, 1995.
[10] V. Parsa, P. A. Parker, and R. N. Scott, “Adaptive stimulus artifact re-
duction in noncortical somatosensory evoked potential studies,” IEEE
Trans. Biomed. Eng., vol. 45, pp. 165–179, Feb. 1998.
[11] , “Convergence characteristics of two algorithms in nonlinear stim-
ulus artifact cancellation for electrically evoked potential enhancement,”
Med. Biol. Eng. Comput., vol. 36, pp. 1–13, 1998.
[12] W. G. Stevens, “The current-voltage relationship in human skin,” Med.
Electron. Biol. Eng., vol. 1, pp. 389–399, 1963.
[13] M. Yuwaraj, A. Gluzmann, P. Madsen, and H. Kunov, “Adaptive
stimulus artifact cancellation in otoacoustic emission testing,” in Proc.
20th Canadian Medical and Biological Engineering Conf., 1994, pp.
106–107.
[14] R. Grieve, P. A. Parker, and B. Hudgins, “Adaptive stimulus artifact can-
cellation in biological signals using neural networks,” presented at the
17th IEEE EMBES Conf., Montreal, P.Q., Canada, Sept. 1995.
[15] S. Haykin, Neural Networks: A Comprehensive Foundation. New
York: Macmillan College, 1994.
Richard Grieve received the B.Sc. and M.Sc.
degrees, both in electrical engineering, in the field
of biomedical engineering, from the University of
New Brunswick, Fredericton, NB, Canada, in 1994
and 1996, respectively.
He is currently employed at Newbridge Networks
in Kanata Ont., Canada. His research interests
include neural networks and their application to
DSP, and the use of such techniques in biomedical
signal processing.
395
Philip A. Parker (S’70–M’73–SM’86) received the
B.Sc. degree in electrical engineering from the Uni-
versity of New Brunswick (UNB), Fredericton, NB,
Canada, in 1964, the M.Sc. degree from the Univer-
sity of St. Andrews, St. Andrews, U.K., in 1966, and
the Ph.D. degree from the UNB in 1975.
In 1996, he joined the National Research Council
of Canada as a Communications Officer and the fol-
lowing year he joined the Institute of Biomedical En-
gineering, UNB, as a Research Associate. In 1976, he
was appointed to the Department of Electrical Engi-
neering, UNB, and currently holds the rank of Professor in that department. He
is also a Research Consultant to the Institute of Biomedical Engineering, UNB.
His research interests are primarily in the area of biological signal processing.
Bernard Hudgins (M’97) received the Ph.D. degree
from the University of New Brunswick (UNB), Fred-
ericton, NB, Canada, in 1991.
He is currently a Senior Research Associate with
the Institute of Biomedical Engineering at UNB
His primary research interests are in the area of
myoelectric signal processing for the control of
artificial limbs.
Dr. Hudgins was the recipient of a Whitaker
Foundation Investigator Award and recently spent
two years on a NATO workgroup assessing alterna-
tive technologies for cockpit applications. He has been the Region 7 (Canada)
representative on the IEEE EMBS Advisory Committee for the past two years.
Kevin Englehart (S’90–M’99) received the B.Sc.
degree in electrical engineering and the M.Sc and
Ph.D. degrees from the University of New Brunswick
(UNB), in 1989, 1992, and 1998, respectively.
He is currently an Assistant Professor of Electrical
and Computer Engineering, and a Research Con-
sultant to the Institute of Biomedical Engineering at
UNB. His research interests include neuromuscular
modeling and biological signal processing using
adaptive systems, pattern recognition, and time-fre-
quency analysis.
Dr. Englehart is a Registered Professional Engineer, and a member of the
IEEE Engineering in Medicine and Biology Society, the IEEE Computer So-
ciety, and the Canadian Medical and Biological Engineering Society.