Journal of Neuroscience Methods 109 (2001) 137– 145
www.elsevier.com/locate/jneumeth
Stimulus artifact removal using a software-based two-stage peak
detection algorithm
Derek T. O’Keeffe a,*, Gerard M. Lyons a, Alan E. Donnelly b, Ciaran A. Byrne b
a
Biomedical Electronics Laboratory, Department of Electronic and Computer Engineering, Uni6ersity of Limerick, Limerick, Ireland
b
Department of Physical Education and Sports Science, Uni6ersity of Limerick, Limerick, Ireland
Received 20 March 2001; received in revised form 30 May 2001; accepted 31 May 2001
Abstract
The analysis of stimulus evoked neuromuscular potentials or m-waves is a useful technique for improved feedback control in
functional electrical stimulation systems. Usually, however, these signals are contaminated by stimulus artifact. A novel software
technique, which uses a two-stage peak detection algorithm, has been developed to remove the unwanted artifact from the
recorded signal. The advantage of the technique is that it can be used on all stimulation artifact-contaminated electroneurophysiologic data provided that the artifact and the biopotential are non-overlapping. The technique does not require any estimation
of the stimulus artifact shape or duration. With the developed technique, it is not necessary to record a pure artifact signal for
template estimation, a process that can increase the complexity of experimentation. The technique also does not require the
recording of any external hardware synchronisation pulses. The method avoids the use of analogue or digital filtering techniques,
which endeavour to remove certain high frequency components of the artifact signal, but invariably have difficulty, resulting in
the removal of frequencies in the same spectrum as the m-wave. With the new technique the signal is sampled at a high frequency
to ensure optimum fidelity. Instrumentation saturation effects due to the artifact can be avoided with careful electrode placement.
The technique was fully tested with a wide variety of electrical stimulation parameters (frequency and pulse width) applied to the
common peroneal nerve to elicit contraction in the tibialis anterior. The program was also developed to allow batch processing
of multiple files, using closed loop feedback correction. The two-stage peak detection artifact removal algorithm is demonstrated
as an efficient post-processing technique for acquiring artifact free m-waves. © 2001 Elsevier Science B.V. All rights reserved.
Keywords: Stimulus artifact; Artifact removal; Digital signal processing; M-wave; Electrical stimulation; EMG blanking; Electrophysiology
1. Introduction
The recording and interpretation of myoelectric activity is a useful technique for understanding muscle
activity for diagnostic, prognostic and therapeutic purposes (Solomonow et al., 1985). The electrically evoked
myoelectric signals (m-waves) of paralysed muscles
stimulated by surface electrodes is also of interest if the
relationship between the electrical activity and the force
within the stimulated muscle is to be investigated (Minzly et al., 1993). However, m-waves may have a stimulation artifact associated with them. This is true
especially when surface electrodes are used to stimulate
the tissue and to detect the resulting generated biopo* Corresponding author. Tel.: + 353-61-202443.
E-mail address: derek.okeeffe@ul.ie (D.T. O’Keeffe).
tentials. While experimental techniques have been developed which minimise the effect of stimulus artifact
contamination, these techniques invariably leave residual artifact, which may interfere with analysis of the
m-wave data (Parsa et al., 1998).
The stimulus artifact is typically spike-shaped followed by an exponential decay (Fig. 1) whose amplitude and time constant is dependent on several factors
(McGill et al., 1982). These factors include experimental technique, stimulator output stage, amplifier design,
electrode orientation, stimulation type and electrode –
skin interface. If the recording site is sufficiently far
away from the stimulus location, then the artifact and
the evoked physiological response will not overlap
(Harding, 1991). The distance between the two sites will
affect the conduction latency period (Fig. 1). A long,
very slow exponential decay may offset the stimulus
0165-0270/01/$ - see front matter © 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 2 7 0 ( 0 1 ) 0 0 4 0 7 - 1
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
138
artifact waveform during data recording, but this can
be eliminated by avoiding amplifier saturation (Harding, 1991; Erfanian et al., 1998).
The stimulus artifact has been modelled by Scott et
al. (1997) as the superposition of three components.
The first is the voltage gradient that appears across the
recording electrodes resulting from the stimulus current
travelling through the limb. The second component is
due to imperfect stimulus isolation that gives rise to a
stray capacitance between the stimulating electrodes
and ground, which results in a second current that once
again creates a voltage gradient at the recording
electrodes.
The third component is due to electromagnetic coupling between the stimulating and recording leads. This
component’s contribution is highly dependent on the
impedance of the recording electrodes, quality of the
shielding on the leads and their position in space.
2. Artifact removal
An analysis of the literature identified many approaches for the removal of stimulation artifact from
muscle (EMG), nerve (ENG) and cerebral (EEG) potentials evoked by electrical stimulation. Since EMG/
ENG/EEG are all fundamentally electrophysiological
signals, a technique developed to remove stimulus artifact contamination from any of the biopotential recordings, may be modified and used to remove artifact from
the other types. These techniques may reduce the effect
of stimulus artifact and improve the fidelity of the
recorded m-wave. However, most techniques suffer
from an inability to adapt to the dynamic nature of
stimulation artifact, due to the non-linearities of the
stimulation procedure and hence suffer residual artifact.
The total removal of the stimulus artifact is necessary
for accurate interpretation of the m-wave (Parsa et al.,
1998). A brief review of the literature follows.
2.1. Hardware techniques
2.1.1. Blanking
A popular method for stimulus artifact suppression is
hardware blanking (Roby and Lettich, 1975 (EEG);
Knaflitz and Merletti, 1988 (EMG)). One variation of
this method is the sample and hold design. These
circuits sample the input signal and switch to the hold
mode during the artifact period (Freeman, 1971 (EEG);
Babb et al., 1978 (EEG)). Hardware-based blanking
circuits are usually triggered by an external synchronous pulse from the stimulator or as proposed by
Minzly et al. (1993) (EMG), they can be triggered from
the stimulus pulse itself.
A major disadvantage of the hardware blanking technologies developed is that the blanking time interval is
usually fixed by a potentiometer. This may lead to
clipping or blanking out of subsequent m-waves if the
blanking interval is set too long or if the stimulation
parameters (frequency/pulse width/amplitude) change.
The approach also fails to take into account the dynamic effect of stimulation whereby the tail of the
stimulus artifact may survive the blanking pulse (if set
too short) and be recorded along with the m-wave.
2.1.2. Stimulator output stage
A combination of the two types of stimulator output
stages was proposed by Pozo and Delgado (1978)
(EEG). It combined the clinical consistency (as the
current delivered is independent of the impedance of
the load) of a constant current output stage during
stimulation with the low impedance advantage of a
constant voltage output stage between stimulus. This
had the effect of minimising stimuli artifact transients.
2.1.3. Filtering
Analogue filtering was used by Solomonow et al.
(1985) (EMG) who implemented an eighth order
Chebyshev low pass filter at 550 Hz. While successful in
eliminating high frequency components of the artifact,
the method allows low frequency components of the
artifact to be passed. This predicament was observed by
Winchman (2000) (EEG) who noted that in the frequency domain, it is very difficult to remove the stimulus artifact using conventional filters, as there exists a
significant overlap between the spectral components of
the neuronal signal and the artifact.
Fig. 1. Contaminated data (stimulus artifact + m-wave).
2.1.4. Amplifier gain
Another artifact removal method was suggested by
Roskar and Roskar (1983) (EMG) who used a digitally
controlled amplifier gain to suppress the stimulus arti-
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
fact signal. A gain of ×1 was used during stimuli
changing to a gain of × 1000 in between stimuli.
However, the authors discussed only qualitative results
and it has been suggested that a gain ratio of 1000 may
not always be sufficient to remove the stimulation
artifact (Knaflitz and Merletti, 1988) (EMG).
A complete stimulation detection system encompassing many hardware design improvements (including
slew rate limiting and windowing) and good experimental technique was proposed by Knaflitz and Merletti
(1988) (EMG). Using the system, the authors were able
to achieve stimulus artifact levels lower than the ambient noise level.
The main problem with any of the hardware approaches described is the inability of the methods to
dynamically adapt to changes in stimulus artifact
shape.
2.2. Software techniques
Following digitisation and storage on a PC, the
possibility of artifact removal during post-processing of
the data exists (Hines et al., 1996 (EMG)). McGill et al.
(1982) (ENG) outlined a straightforward software technique to remove the stimulus artifact. The principle
involved acquiring an estimate of the stimulus artifact
waveform and subtracting this from the contaminated
signal.
Three methods of recording an uncontaminated stimulation artifact waveform were outlined:
(1) Sub-threshold stimulation: This is where the stimulus intensity is reduced below the threshold for nerve
excitation.
(2) Recording off ner6e: This method uses a second
pair of recording electrodes positioned away from the
nerve under examination to record a purely artifactual
signal.
(3) Double stimulus method: This is where a second
stimulus pulse is applied to the nerve during its refractory period. During this refractory period the stimulus
fails to evoke a nerve response allowing the recording
of the artifact signal.
Once an estimated stimulation artifact signal was
acquired, it was subtracted from the recorded m-wave
with artifact contamination, leaving an estimated uncontaminated m-wave (McGill et al., 1982).
Kiss and Shizgal (1989) (EEG) used the digital subtraction method to remove stimulation artifact from
recorded compound action potentials. They were able
to remove most of the stimulation artifact but acknowledged that the method rarely removes the artifact entirely. They also pointed out that the subtraction
method of artifact removal degrades the signal-to-noise
ratio of the data. Blogg and Reid (1990) (EMG)
recorded myoelectric signals and digitised the data using a 12-bit ADC with 10 kHz sampling. The method
139
involved recording a sub-threshold artifact as a template and also recording the normal contaminated action potentials. Using linear interpolation, amplitude
scaling and time synchronisation, the stimulus artifact
was removed by subtraction. However, Blogg and Reid
(1990) reported that their synchronisation with the
stimulus pulse spike was poor (Merletti et al., 1992).
The subtraction method of artifact removal suffers
from residual artifact remaining at the end of processing. This can be attributed to the fact that the ‘artifact
template’ that is subtracted from the contaminated
signal is taken usually once at the beginning of the
experiment and thus does not take into account the
time varying nature of the stimulus artifact and non-linearities present in stimulus artifact generation and
recording (Stephens, 1963; McGill et al., 1982).
Estimates of the average stimulus artifact waveform
was investigated by Winchman (2000) (EEG) who used
digital averaging techniques to create a more representative stimulus artifact template. However, even though
the method is significantly better in constructing a more
valid stimulus artifact, it is acknowledged by Winchman (2000) that the method depends on the assumption
that the shape of the stimulus artifact remains the same
throughout repeated stimulation.
A solution to the problem of the non-linearity in the
stimulus component was proposed by Parsa et al.
(1998) (ENG). Again the method involved recording
either a sub-threshold or dual recording of an uncontaminated stimulus artifact signal and subtracting this
from the contaminated m-wave. By using Volterra series expansion, they were able to model the non-linearities of the recorded signal and by using two different
non-linear adaptive filter techniques. They were able to
achieve better results in artifact suppression than other
techniques that took a single sample of the stimulus
artifact shape and used this waveform for all subsequent processing even though their are non-linearities
in stimulus artifact generation (Stephens, 1963; McGill
et al., 1982).
Harding (1991) (EEG) proposed another method of
artifact removal using artifact peak detection and mathematical equation fitting of the artifact waveform. The
method relies upon several assumptions about the artifacts and physiological signal. Again it is acknowledged
that residual artifact remains in the data as the mathematical equations do not fit the data perfectly. The
correct operation of the method also depends on a
noise free data matrix, which is not always possible in
experimental situations.
Any attempt to use hardware or software blanking
techniques on electrophysiological signals contaminated
with stimulus artifact depends on the assumption that
the stimulus artifact and the biopotential are not overlapping. With short distances between the recording site
and the stimulation source or pulse width modulation
140
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
involving long pulses the artifact and the biopotential
will overlap and another approach is needed (Harding,
1991 (EEG); Erfanian et al., 1998 (EMG)).
This paper proposes a new approach for the dynamic
removal or stimulus artifact based on a fundamental
characteristic of the recorded signal. When recording an
m-wave, the recorded artifact signal is usually one to
two orders of magnitude greater than the pure m-wave
(Knaflitz and Merletti, 1988). The method also assumes
that the m-wave and artifact are non-overlapping (as
discussed) which is true if the latency of the evoked
response from the stimulus pulse is sufficiently long
(Harding, 1991 (EEG)). Using a two-stage peak detection algorithm, the data matrix is scanned until a peak
is detected and using two levels of amplitude threshold,
the entire stimulus artifact is removed leaving the uncontaminated m-wave remaining.
3. Method
3.1. Algorithm
The technique developed for stimulus artifact removal is a software-based solution. The artifact removal program was written in MATLAB (Mathworks,
Nantick, MA). The program removes the stimulation
artifact, by setting high (HT) and low (LT) threshold
values (Fig. 1) to identify, isolate and remove the
stimulus artifact, while leaving the m-wave intact. The
algorithm removes the total artifact including both
positive and negative spikes and any exponentially decaying tail (Fig. 2). The solution is flexible enough to
work for any file size and for any type of recorded
stimulus contaminated electroneurophysiologic data.
3.2. Program operation
The program was designed to scan for negative as
well as positive stimulation artifacts, which it does by
considering the absolute magnitude of the data points.
The program control flow is as follows:
(1) The recorded contaminated (stimulus artifact+
m-wave) data is inputted to the program.
(2) The program initially calculates and removes any
offset in the recorded signal. It then scans through the
absolute data matrix of the artifact-contaminated signal
and calculates the maximum value in the array. This
value corresponds to the peak artifact amplitude. Then
the program initially sets two threshold values: a high
threshold (HT) which corresponds to the maximum
peak detected divided by two and a low threshold (LT)
set as 1/20 times the maximum peak value (Fig. 2). By
using the absolute matrix both positive and negative
polarities are processed with the same HT/LT values.
The program will remove both single and multiphasic
artifact waveform, as it removes each artifact as it
encounters it.
(3) The program then scans the matrix until it
reaches an amplitude value that passes the LT level in
an increasing direction. As this could be the beginning
of a stimulation artifact peak, the program marks the
initial LT index point. The program then monitors the
data to establish if the amplitude continues to increase
and pass the HT value. However, if after passing the
initial LT value, the data decreases and passes the LT
value again, then it is not an artifact and is either an
m-wave or sporadic noise, therefore, the program will
ignore it and continue scanning the matrix for artifacts
(right-hand loop of flow diagram— Fig. 2).
(4) Alternatively, if after passing the initial LT value,
the data series amplitude continues to increase and
passes the HT value, then the program marks this
sequence (which started at the passing of the initial LT
value) as a valid stimulation artifact detect. The program continues to monitor the progress of the data
noting when it passes the HT (decreasing) and then the
LT (decreasing). The program then marks and records
the length of the artifact array sub-set (from the matrix
index of the initial LT crossing (− 0.1 ms) to the matrix
index of the final LT (+ 0.1 ms) crossing with a valid
peak detect condition in-between). The program then
replaces this stimulation artifact sub-set with zeros (or
pre-artifact baseline). 0.1 ms was chosen on either side
of the artifact deletion window (LT2–LT1) as it is one
sample period (10 kHz sampling) and is necessary to
ensure proper functioning of the programs nested loops
(left-hand loop of flow diagram— Fig. 2).
It is important to note that if the index markers from
the first LT to the first HT were replaced with zeros,
then only the stimulation artifact spike would be removed. However, it is also necessary to remove any
exponential decay of the artifact signal. Therefore, the
program continues scanning the matrix and monitors
the progress of the data series until the stimulation
artifact amplitude drops back to below both the HT
and the LT levels (Fig. 1).
(5) When developing the technique for batch processing of multiple files, a method of checking the effectiveness of artifact removal was included (Step 5 of flow
diagram— Fig. 2). M-wave data files from recordings
were given coded filenames e.g. CB10H100. ‘CB’ signifies the subject’s initials, ‘10H’ signifies the stimulation frequency (10 Hz) and ‘100’ signifies the
stimulation pulse width. The stimulation was applied
for 10 s with a duty cycle of 2:3 and the type of
stimulation waveform was asymmetric biphasic. Following visual inspection of multiple artifact files, it was
observed that the recorded stimulation artifact typically
had three (repolarising factor) spikes for every stimulation artifact (due to repolarising effects). Using this
empirically derived information a formula was devel-
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
141
Fig. 2. Algorithm flow diagram.
oped to find the approximate number of stimulation
artifacts in any recorded file. The formula developed to
predict the number of stimulation artifact peaks (SAPP)
is:
SAPP =(A)×(B)× (C) ×
D
,
E
(1)
where A is the duty cycle on time; B is the stimulation
frequency; C, repolarising factor (dependent on stimulation waveform type); D, data recording time (multiples of 5 s); and E, sum of duty cycle on and off period.
Using the above example for 10 s of recorded data,
the program would expect 2×10 ×3 ×2 = 120 stimulation artifact peaks (40 stimulation artifacts). Since the
predicted number of artifact peaks is not an absolute, if
the number of artifacts detected was within 5% of that
expected using Eq. (1), then the artifact detection process was deemed successful. The 5% tolerance was
developed empirically, after examining the output of
the recorded data files whose number of artifacts varied
from the formula predicted result by less than or equal
to 5%. When the algorithm loads the data file into
memory it strips the filename to extract the required
information. Using this predicted value for the number
of peaks in the data matrix, the program compared this
figure with the number of artifact pulses it had isolated.
If the number of stimulation artifacts detected is outside of this 5% tolerance, then the program automati-
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
142
cally changes the values of the HT and LT appropriately and scans the matrix again, performing accuracy
iterations as follows:
If the number of artifact pulses removed is too high
( \ 5% of the predicted value), the program increments
(by 1%) the HT/LT values and re-runs the algorithm, it
continues to increase these threshold values if the number of artifact pulses removed is still too high. Reciprocally if the number of artifact pulses removed is too low
( \ 5% of the predicted value) the program decrements
(by 1%) the HT/LT value until the predicted value of
stimulus pulses matches the actual value of stimulus
pulses identified. The program changes the value of the
HT and the LT by varying the denominator value of
the equations:
High threshold=
Low threshold=
Peak amplitude
,
X
Peak amplitude
,
Y
where X is initially set as 2, but varies according to
feedback iterations; Y is initially set as 20, but varies
according to feedback iterations.
It continues to do this iteratively until the number of
expected artifact waveforms (within 5%) match the
number of artifact waveforms removed. To avoid endless program loops and potential loss of m-wave data,
the program stops the iterations before the expected
number of artifact peaks (within 5% tolerance)=number of artifact peaks removed depending on two
factors:
(1) The program allows a maximum of 20 accuracy
feedback iterations. Most data files find the predicted
number of artifacts (within 5%) in less than 10
iterations.
(2) It is also important that the program does not
keep lowering its thresholds, potentially removing mwaves, for the sake of meeting the predicted target of
stimulation peaks. Therefore as an error control, the
program calculates the average maximum of the clean
m-waves and never allows the HT value to drop below
this amplitude. This safeguards the important m-wave
data from closed loop feedback iterations, while max-
imising the efficiency of the program to remove the
artifact pulses.
Having safeguards built into the program code allows for batch processing of multiple files without the
need to check each individual file. The program automatically generates output graphs and also generates an
operation log file (Fig. 3), both of which give the user
feedback information on the programs operation
(Note: lo – th= low threshold; hi – th= high threshold).
In Fig. 3, the program output file shows that it expected
40 (9 , less than, or equal to 5%) stimulation artifact
waveforms. Therefore, it would accept 38–42 artifacts
detected as a valid result. After setting up the high and
low thresholds with initial conditions, the first iteration
found 26 stimulation artifact waveforms. Over the next
few iterations, the program continued to lower its
threshold values until, iteration four, when it found all
40 of the 40 stimulation artifact waveforms. It is important to note that the program will always try to achieve
40/40 artifacts removed, but will accept a result within
5% of predicted. However, this does not mean that the
program reaches 5% of the predicted value and stops
looking, it always tries to find the maximum number of
artifact peaks without breaking the error control rules
outlined above.
4. Data collection
As discussed, several techniques exist to minimise the
effect of stimulus artifact contamination during collection of m-wave data (McGill et al., 1982; Scott et al.,
1997; Erfanian et al., 1998). These techniques were
incorporated into the experimental design. After the
researchers received ethical approval, ten subjects provided informed consent to participate in a study to test
the effectiveness of the detection algorithm. The subjects were electrically stimulated at the common peroneal nerve to elicit tibialis anterior contraction. The
subjects were seated in a CON-TREX MJ machine; a
rotary biomechanical test and training system (CONTREX, CMV-AG Zurich, Switzerland) in order to
measure the force produced from the contraction of the
tibialis anterior, for future m-wave-torque research.
Fig. 3. Output of program process log file.
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
A commercial programmable stimulator (BMR
NeuroTech Inc., 1995, Galway, Ireland) was used to
provide stimulation pulses to two commercial 32 mm
diameter woven conductive cloth stimulation electrodes
(Nidd Valley ACUPAD, Yorkshire, England). Strategic
placement of the stimulation electrodes with respect to
the recording electrodes is critical in minimising the
stimulus artifact and obtaining a good signal (Erfanian
et al., 1998). The area to be stimulated was clipped (not
shaved) of hairs if necessary to ensure a better surface
contact.
Stimulation electrodes were positioned to elicit good
dorsiflexion at 35 Hz 300 ms 50 mA. The stimulating
electrodes were placed over the common peroneal nerve
(1 cm below head of fibula) and the tibialis anterior
muscle. A constant voltage stimulator was used to
ensure the artifact transients were relatively short
(Knaflitz and Merletti, 1988). An asymmetric biphasic
stimulation waveform was used as a biphasic waveform
accelerates the discharge of tissue and electrodes capacitance’s through the stimulator output stage (Spencer,
1981). During the first phase the stimulation, current
rises rapidly to a positive peak and then reduces in a
smooth exponential fall towards zero. In the second
phase, the current reverses direction and slowly reduces
from a negative peak towards zero again. Overall the
net current flow is approximately zero, since the charge
transferred in each phase is equal but opposite.
EMG recording electrodes were positioned on the
midline of the muscle belly 6 cm distally from the
stimulation electrodes (McGill et al., 1982). The limb
was grounded between the stimulating and recording
sites to reduce 50 Hz interference, to hold the mean
voltage of the limb near ground and to prevent
transthoracic current flow if there is a fault in the
stimulator (McGill et al., 1982). The EMG interelectrode spacing chosen was 20 mm, with an electrode
diameter of 5 mm. The electrodes were fixed in a
pre-amplifier unit with dimensions of 30 mm× 20
mm × 6 mm. The Ag – AgCl electrodes protrude 0.5
mm from the unit. Good electrical contact was achieved
between the recording amplifier and the muscle under
investigation by shaving the recording site and by
mildly abrading and cleaning the skin. Electrode gel
was also used to aid electrical conductivity and to
minimise skin impedance variation effects (McGill et
al., 1982). The electrodes were fixed to the limb using
Velcro straps to minimise motion artifact.
The EMG amplifier (Jim Clark Ltd., Newcastle, England) incorporated several features to make the device
less sensitive to stimulus artifact. The amplifier chosen
had a large input impedance (\10 MV). This allows
the amplifier to operate independently of changes in the
skin impedance. Since there is a large stimulation common mode artifact and power line interference, it was
important for the amplifier to have a high (\ 100 dB)
143
common mode rejection ratio (deLuca, 1997). The amplifier was internally clamped with diodes to protect it
from damage due to the presence of stimulus artifact
and it also had a fast settling time so it could recover
quickly from saturation. The amplifier gain was set to
1000 for the experiment.
The m-wave data was recorded with a MACLAB (AD
Instruments Ltd., CA) data acquisition card. It was
necessary to ensure that the full artifact signal was
digitised, so the signal was initially sampled at 40 kHz.
Spectral analysis of the data showed that the m-wave
plus artifact signal had no significant spectral components above 4 kHz (Winchman, 2000). Therefore, a
lower sampling rate could be chosen to ease processing
overhead and data file size while still maintaining signal
integrity. Subsequently data were sampled at 10 kHz,
with an anti-alias filter at 5 kHz. Ten subjects were
stimulated at three different stimulation frequencies (10,
20 and 30 Hz) and seven different pulse widths (100,
150, 200, 250, 300, 350 and 400 ms). This generated 210
data files with a wide variety of recorded data.
5. Results
Rigorous testing of the program algorithm was carried out using the collected experimental data.
5.1. Manual testing
If the program algorithm processes a single data file,
the user can change the threshold levels manually using
visual feedback of the results to achieve 100% stimulation artifact removal. However, in situations where the
user visual feedback is removed (e.g. batch processing
of multiple files), the program uses Eq. (1) to predict an
approximate number of stimulation artifacts to be removed. To assess the effectiveness using this method in
batch processing multiple files, a comparison was made
of the number of artifacts removed by the program
versus the number of artifacts expected (as determined
by Eq. (1)). As the predicted number of artifacts
present is not exact, a 5% tolerance figure was applied
to this number and accepted as a valid result. Several
files were processed with the program and then the
output data file was compared with the process log file.
In all cases the program found the predicted number of
artifacts (within 5% tolerance).
5.2. Batch processing
Once the accuracy of the formula feedback technique
was established, 210 data files were batch processed.
The program successfully removed all of the stimulation artifact waveform from all 210 batch processed
files. An example of the stimulation artifact removal
144
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
Fig. 4. (A, B) Waveform stimulated at 10 Hz, pulsewidth = 200 ms. (C, D) waveform stimulated at 30 Hz, pulsewidth =350 ms.
program output can be seen in Fig. 4. Fig. 4((A) and
(C)) shows a stimulation artifact-contaminated m-wave
and Fig. 4((B) and (D)) shows both the contaminated
wave in more detail and the subsequent cleaned wave
after processing.
The program will always try to find the exact predicted number of artifact waveforms, but if after trying
a pre-set number of iterations and/or threshold violation conditions, it will accept a value of within 5% that
was encountered in the processing.
In some situations 0– 5% greater than the predicted
number of artifacts are removed. This is due to the
limitations of the prediction formula (Eq. (1)). In the
formula of Eq. (1), the repolarising factor is equal to
three spikes (on average), however, in some of the
recorded data files, stimulation artifacts have a combination of four and three spikes. This causes a greater
number of artifacts to be present compared to what is
predicted. The program tolerates this inaccuracy in the
predication model by allowing 101–105 peaks (5%) of a
predicted 100, to produce a successful result.
6. Discussion
The developed MATLAB (Mathworks) program eliminates stimulus artifact from contaminated m-waves.
Unlike other artifact removal approaches, the technique
removes both the stimulus artifact spike and the decaying exponential by requiring the data to pass two
threshold points (LT and HT) on the rising edge and
two threshold points (HT and LT) on the falling edge.
The main advantage of the technique is that by using
a two-stage threshold approach, the technique can dynamically adapt to different stimulation artifact waveforms. When the experimental data was analysed the
D.T. O’Keeffe et al. / Journal of Neuroscience Methods 109 (2001) 137–145
maximum artifact duration was found to be 1.9 ms and
the minimum artifact duration was 0.9 ms. The mean
artifact duration was 1.3 ms with a standard deviation
of 0.25 ms, which is an artifact duration variation of
20%.
For the technique to work properly it is important
that the m-wave and the artifact signal do not overlap,
as the program would remove the artifact from the
initial LT up to the final LT, which may include part of
the m-wave signal. The residual artifact remaining after
removal is negligible, as invariably the LT value is set
(dynamically or manually) to just above pre-artifact
baseline level to avoid continuous triggering of the
algorithm by system noise and to ensure that the maximum amount of contamination is removed.
The user is not required to record any pure artifact
signals for use in post-processing manipulation methods, such as subtraction. The method is also independent of the variation in artifact shape and duration
between stimulation pulses due to non-linearities, as the
program removes each artifact it encounters in the data
file. The stimulus artifact waveform is replaced with
zeros, which makes this technique useful for trials
studying small signal amplitudes contaminated with
stimulus artifact (residual stimulus artifactB10 mV;
Hines et al., 1996).
The program allows batch processing of multiple
files, using feedback controlled threshold correction
based on a developed formula that predicts the number
of stimulation artifacts present. It is, however, acknowledged that the formula was developed empirically from
the data collected in this experiment and may have to
be modified when used as the feedback control mechanism in the processing of other types of data. The
program was initially developed using MATLAB (Mathworks), however, when batch processing was implemented, it was observed that data processing took
longer than expected. Therefore, the program was converted to C code (Borland Inc., Scots Valley, CA) and
compiled as a standalone executable file. This greatly
reduced processing time compared to the command line
interpretation processing of the MATLAB approach (18
times faster, on the same operating system and computer platform).
The technique removed 100% of stimulation artifact
waveforms when tested with experimental data, taken
from ten subjects whose tibialis anterior was electrically
stimulated with various stimulation waveform parameters. At the end of batch processing, the computer
program produces a log file, which summarises the
program operation on each data file. Using this output
file, with automatically generated graphs, the user can
ensure that the program is working properly.
The software algorithm developed is proposed as a
useful tool, in the removal of stimulation artifact waveforms from contaminated electroneurophysiologic data.
145
Acknowledgements
The authors would like to acknowledge the support
of the Irish Mid-Western Health Board in funding this
research.
References
Babb TL, Mariani E, Strain GM, Lieb JP, Soper HV, Crandall PH.
A sample and hold amplifier system for stimulus artifact suppression.
Electroenceph Clin Neurophys 1978;44:528 – 31.
Blogg T, Reid WD. A digital technique for stimulus artefact reduction.
Electroenceph Clin Neurophys 1990;76:557 – 61.
BMR NeuroTech Inc. BMR NeuroTech NT2000 operators manual,
1995.
deLuca CD. The use of surface electromyography in biomechanics. J
Appl Biomech 1997;13:135 – 63.
Erfanian A, Chizeck HJ, Hashemi RM. Using evoked EMG as a
synthetic force sensor of isometric electrically stimulated muscle.
IEEE Trans Biomed Eng 1998;45(2):188 – 202.
Freeman JA. An electronic stimulus artifact suppressor. Electroenceph
Clin Neurophys 1971;31:170 – 2.
Harding GW. A method for eliminating the stimulus artifact from digital
recordings of the direct cortical response. Comp Biomed Res
1991;24:183 – 95.
Hines AE, Crago PE, Chapman GJ, Billian C. Stimulus artifact removal
in EMG from muscles adjacent to stimulated muscles. J Neurosci
Meth 1996;64:55 – 62.
Kiss I, Shizgal P. Improved artifact rejection and isolation of compound
action potentials by means of digital subtraction. J Neurosci Meth
1989;30:219 – 29.
Knaflitz M, Merletti R. Suppression of stimulus artifacts from myoelectric-evoked potential recordings. IEEE Trans Biomed Eng
1988;35(9):758 – 63.
McGill KC, Cummins KL, Dorfman LJ, Berlizot BB, Luetkemeyer K,
Nishmura DG, et al. On the nature and elimination of stimulus
artifact in nerve signals evoked and recorded using surface electrodes. IEEE Trans Biomed Eng 1982;29(2):129 – 35.
Merletti R, Knaflitz M, DeLuca CJ. Electrically evoked myoelectric
signals. Crit Rev Biomed Eng 1992;19(4):293 – 340.
Minzly J, Mizrahi J, Hakim N, Liberson A. Stimulus artefact suppressor
for EMG recording during FES by a constant current stimulator.
Med Biol Eng Comp 1993;31:72 –5.
Parsa V, Parker P, Scott R. Convergence characteristics of two
algorithms in non-linear stimulus artefact cancellation for electrically evoked potential enhancement. Med Biol Eng Comp
1998;36:202 – 14.
Pozo FD, Delgado JMR. Hybrid stimulator for chronic experiments.
IEEE Trans Biomed Eng 1978;25(1):92 – 5.
Roby RJ, Lettich E. A simplified circuit for stimulus artefact suppression. Electroenceph Clin Neurophys 1975;39:85 – 7.
Roskar E, Roskar A. Microcomputer based electromyographic recording system with stimulus artefact suppression. In: Third Medical
Conference on Biomedical Engineering, 1983. Yugoslavia: Portoroz.
Scott RN, McLean L, Parker PA. Stimulus artefact in somatosensory
evoked potential measurement. Med Biol Eng Comp 1997;35:
211 – 5.
Solomonow M, Baratta R, Miwa T, Shoji H, D’Ambrosia R. A
technique for recording the EMG of electrically stimulated skeletal
muscle. J Orthoped 1985;8(4):492 – 5.
Spencer HJ. An automatic, optically isolated, biphasic constant current
stimulator adaptor for artefact suppression. Electroenceph Clin
Neurophys 1981;51:215 – 7.
Stephens WGS. The current – voltage relationship in human skin. Med
Elec Biol Eng 1963;1:389 – 99.
Winchman T. A digital averaging method for the removal of stimulus
artifacts in neurophysiologic experiments. J Neurosci Meth
2000;98:57 – 62.