Journal of Neuroscience Methods 116 (2002) 165 /177

www.elsevier.com/locate/jneumeth

Choosing a wavelet for single-trial EMG

Martha Flanders *
Department of Neuroscience, University of Minnesota, 6-145 Jackson Hall, Minneapolis, MN 55455, USA

Received 7 January 2002; received in revised form 25 February 2002; accepted 25 February 2002

Abstract

A wavelet analysis was developed to measure the timing of multiunit bursts in surface electromyograms (EMGs) from single trials.
EMG data were taken from eleven elbow and/or shoulder muscles during reaching movements in six different directions, at a range
of speeds. A relatively simple wavelet (db2) was chosen, and the analysis focused on wavelet coefficients at an intermediate scale
(D3), where the wavelet length approximately matched the wavelengths present in EMG bursts. Burst times were identified from the
peaks of the coefficient traces and were plotted as a function of movement time. Linear regression revealed significant relations in
most cases, and thus served to validate the wavelet burst identification. With a few exceptions, burst timing scaled in a manner
approximately similar to the scaling of movement time. As shown previously with different analytical methods, both within and
across joints, EMG bursts were not confined to distinct ‘agonist’ and ‘antagonist’ time frames, but instead showed a variety of
phases relative to speed or joint torque. # 2002 Elsevier Science B.V. All rights reserved.

Keywords: Reaching; Electromyography; Motor pattern generation; Triphasic pattern; Event detection; Daubechies

1. Introduction been used to quantify EMG patterns. For example, our

One of the most amazing functions of the nervous
system is to coordinate the activities of the many motor
units involved in limb movement. Our understanding of
this phenomenon is largely derived from studies of the
cat or frog hindlimb and studies of the primate arm.
Early research on arm movement established that fast,
targeted movements are subserved by three sequential
bursts of muscle activity. The ‘first agonist burst’
provides a propulsive force to accelerate the limb, the
‘antagonist burst’ provides a braking force, and then a
smaller ‘second agonist burst’ helps to end the move-
ment. The entire three-burst pattern is sometimes
referred to as ‘triphasic’ (e.g. Angel, 1974; Ghez and
Martin, 1982).
The primary source of information in studies of
motor patterns is the electromyogram (EMG). A
recording taken directly from the active muscles, EMG
can be used to recognize the firing of a single motor unit
(SMU), or (more often) to estimate the intensity and
timing of multiunit activity. A variety of methods have
* Tel.:
/1-612-624-6601; fax:
/1-612-626-5009.
E-mail address: fland001@umn.edu (M. Flanders).
0165-0270/02/$ - see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 2 7 0 ( 0 2 ) 0 0 0 3 8 - 9
group has developed a procedure to subtract away the
component of the EMG signal that is related to anti-
gravity forces, in order to isolate and measure the phasic
portion of the signal (Flanders and Herrmann, 1992;
Buneo et al., 1994). However, some aspects of the timing
of phasic activation remain elusive. It has gradually
become clear that the triphasic pattern mentioned above
is an oversimplification in that not all active muscles fire
exactly in phase with the main agonist or antagonist
(Hoffman and Strick, 1999). In a classic study of
reaching in a horizontal plane, Wadman et al. (1980)
showed that the relative timing of shoulder and elbow
muscles changed as a function of movement direction
(see also Karst and Hasan, 1991). For reaching in
vertical planes, Flanders et al. (1996) showed that the
movements are often associated with an even more
asynchronous pattern, which sweeps across the ‘ago-
nists’ and ‘antagonists’ within each joint.
If the temporal pattern of muscle activation can be
established, should we expect this pattern to simply scale
in time to subserve faster and slower movements?
Several studies (Soechting and Lacquaniti, 1981; Atke-
son and Hollerbach, 1985; Nishikawa et al., 1999) have
shown that joint rotations and hand speed profiles scale
166 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
in amplitude and in time, to create a ‘family’ of reaching
movements between a particular pair of targets, at a
wide range of peak speeds. For this same ‘family’ of
movements, Hollerbach and Flash (1982), Atkeson and
Hollerbach (1985) also showed that the dynamic com-
ponent of joint torque scales in amplitude and time,
whereas the anti-gravity component scales only in time.
Our initial study of time scaling suggested that the
EMG time base scales in the same manner as joint
torques (Flanders and Herrmann, 1992), but a subse-
quent study suggested that time scaling (i.e. the slope of
the relation between EMG time base and movement
time) was different for different muscles (Buneo et al.,
1994). Specifically, we found a tendency for the time
base of shoulder muscles to scale more than the time
base of elbow muscles, and for the time base of flexors
to scale more than the time base of extensors. Because
we used only one movement direction in this study
(upward and forward from a standard initial posture) it
was not clear whether this time-scaling difference
represented a difference between agonists and antago-
nists, a difference between shoulder muscles and elbow
muscles, or a muscle-specific pattern, perhaps related to
differences in musculoskeletal mechanics (such as differ-
ences in series elastic compliance; Zajac, 1989; Yama-
guchi et al., 1990; Soechting and Flanders, 1997).
Furthermore, this result might have been related to the
analytic technique of scaling and fitting the entire
waveform of averaged EMG signals. Therefore the
present study was designed to revisit this issue using
more movement directions, more muscles, and an
entirely different analytic technique. Specifically, we
wished to avoid problems inherent in averaging data
from trials with non-identical movement times. We
therefore developed a method for measuring burst
timing in data from single trials.
2. Methods
2.1. Experimental design
We used small bipolar surface electrodes (2 mm
diameter, 2 cm apart) to record from eleven superficial
arm muscles as human subjects made reaching move-
ments at a range of speeds. The three subjects (two 5?9ƒ
females, subjects A and B, and one 6?0ƒ male, subject C)
gave informed consent and then stood in a standard
initial posture with the upper arm vertical and the
forearm horizontal, in a parasagittal plane. The forearm
was neutral with regard to pronation/supination and the
subject held a pen with a sphere 3.5 cm in diameter (a
ping-pong ball) glued to the tip. The ball was used to
contact two other balls placed on sensitive switches at
the locations of the initial and final targets. The initial
target always corresponded to the hand location in the
standard initial posture, and the final target was placed
in six different locations in different blocks of trials.
Each final target was 35 cm away, and in one of six
directions from the initial target location. These direc-
tions were chosen to cover the entire workspace and
were named as follows: (1) ‘out’ was directed 458
upward from the horizontal plane and 458 medial
from the parasagittal plane of the initial target; (2)
‘back’ was down and back from the initial target, in line
with ‘out,’ in the opposite direction; (3) ‘up’ was straight
upward; (4) ‘down’ was opposite to ‘up’; (5) ‘left’ was to
the subject’s left, in the horizontal and frontal planes of
the initial target; and (6) ‘right’ was opposite to ‘left.’
Movement onset and end were detected by switches.
The switches were arranged so that the reaching move-
ments would resemble those studied in previous experi-
ments, without unusual accuracy demands or
constraints due to making contact with the target. Since
this technique differs from that of previous studies
(where movement beginning and end were generally
measured by scoring a speed profile) we also conducted
a preliminary experiment in which movements between
the switches were recorded with a magnetic coil system
(Polhemus, 3space Fastrak). Fig. 1 shows the corre-
spondence between the averaged hand speed profile and
the average time of triggering of the initial and final
switches (dashed lines) for subject A. Prior to averaging,
the speed profiles were smoothed with a two-sided
exponential filter. Each panel of this Figure represents
Fig. 1. The hand speed profile scales in amplitude and time to produce
a ‘family’ of reaching movements with different movement times. In
each panel, the speed profile is the average of data from 3 to 5 trials
within a 100 ms range of movement times. These profiles represent the
tangential velocity of a marker on the wrist of subject A, during
outward reaching movements. Movement times for each trial were
defined as the time between the closing of a switch at the starting
location (switch #1) and the opening of the switch at the target
location (switch #2). Averages were aligned on the time of switch #1
(earliest dashed lines). The average time of the second switch is shown
by the later dashed lines, and the standard deviations are shown by
error bars.
 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177 167

averages of 3 /5 trials, with movement times spanning a with averaged waveforms, and to test previous results
particular 100 ms range. using a qualitatively different analytical approach. We
During the experiment, movement direction was therefore developed a technique that allowed quantifica-
organized into 8-trial blocks and the blocks were tion of discrete times of EMG bursts, in records from
repeated six times (in random order) for a total of 288 individual trials. We then used a regression of burst time
trials (6 directions /6 repeats/8 trials). Within each vs. movement time to combine the results across trials.
block, subjects were coached to move at a range of This approach allowed us to view time base scaling
speeds by announcing each trial as ‘fast,’ ‘medium,’ across movement times (using regression slopes) and
‘slow,’ or ‘very slow’ (two times each). burst timing for a given movement time (using regres-
sion intercepts).
2.2. Electromyography The surface EMG signal has a baseline voltage of
zero. Under our recording conditions, the action poten-
During each recording session surface EMG signals tial of a single motor unit (SMU) spanned about 10/20
were taken from brachioradialis and brachialis (elbow ms; it could sometimes be recognized as a characteristic
flexors), biceps (elbow and shoulder flexor), pectoralis waveform occurring periodically, at intervals of about
major and anterior deltoid (shoulder flexors), medial 100 ms (an example is marked with asterisks in Fig. 2A).
deltoid (shoulder abductor), posterior deltoid and In our previous studies of single motor units, we used
latissimus dorsi (shoulder extensors), long head of intramuscular fine-wire electrodes and used the true
triceps (shoulder and elbow extensor), and lateral triceps
motor unit waveform as a cross-correlation template to
and medial triceps (elbow extensors). Although our
locate the time of each SMU potential (Herrmann and
electrodes were small, it is impossible to rule out cross-
Flanders, 1998). In an analogous fashion, we set out to
talk, and we assume that each electrode was most
use an appropriate wavelet as a cross-correlation
sensitive to the motor units directly under it and
decrementally less sensitive to distant motor units.
Evidence from studies of single motor units suggests
that the activation of deltoid and biceps motor units is
organized along a continuum (Herrmann and Flanders,
1998). Thus for biceps and pectoralis major we placed
our electrodes centrally, instead of attempting to record
separately from each of the two heads. EMG signals
were amplified and filtered according to standard
methods (Flanders et al., 1996). The data were digitized
at 1000 Hz so that each successive sample was separated
by 1 ms.

2.3. Wavelet analysis

In the following sections, we explain the rationale for

using wavelet analysis on EMG signals and we describe
the technique. After discussing the background reason-
ing, we will show how a particular wavelet (db2/D3) can
be used as a filter, to recognize the time of occurrence of
a multiunit burst. Following this example, we will
present a more detailed description of the technique,
along with further rationale for choosing this technique
and this particular wavelet.

2.3.1. Rationale for wavelet analysis

The analysis of EMG timing is generally performed
on the average waveform from 5/10 identical move-
ments (i.e. same direction, same distance, same speed).
However, in this experiment we used six movement
directions and many speeds, and therefore it was
impossible to gather a large number of ‘identical’ trials
within a single recording session. Furthermore, we
hoped to avoid some of the ambiguities associated
Fig. 2. Identification of the time of an EMG burst using wavelet
coefficients. In panel A, an example of an EMG burst is taken from
medial triceps for an outward movement in subject A. The Daubechies
2 wavelet (db2) at a middle scale (D3) was convolved with the EMG
trace. In panel B, the resulting wavelet coefficients are shown in full
and downsampled versions. The time of occurrence of the burst was
taken as the peak of the absolute value of the downsampled signal
(circle). In panel C, db2 wavelets of various scales are illustrated.
168 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
template, to detect the time of occurrence of multiunit
bursts. We sought to identify a single wavelet, of a
particular shape and length, which resembled a generic
multiunit burst.
2.3.2. Wavelets as filters and event detectors
A wavelet is a waveform of finite length and an
average value of zero (see inset of Fig. 2A). In general,
wavelet analysis can be viewed as a filtering process:
c(t) Si wi EMG(ti) (1)
where wi is a vector with i elements (representing the
wavelet), EMG is a time series, and c(t) is the product of
the convolution (which gives the wavelet coefficients).
Fig. 2 shows an example of a particular wavelet being
convolved with an EMG trace (Fig. 2A) to yield a time
series of wavelet coefficients (Fig. 2B). Since the wavelet
shape was well correlated with the rapid fluctuations
occurring at the center of the EMG burst, the peak-to-
peak amplitude of the coefficient trace was very high at
this time.
During a typical wavelet analysis, wavelets of differ-
ent lengths are successively applied to a given signal.
This is shown schematically in Fig. 2C. In many wavelet
applications, the portion of the signal that is well
correlated with the wavelet of the shortest length (first
detail, D1) is subtracted away, effectively removing the
high frequency components that typically represent
noise. As explained below, this filtering process is
implemented by applying a discrete series of paired
filters (see Fig. 4B) to repeatedly split the signal into
high frequency (detail, D) and low frequency (approx-
imation, A) bands (see Strang and Nguyen, 1997). To
avoid doubling the number of data points at each stage,
the signal is progressively downsampled, by retaining
only even numbered elements. As illustrated in the lower
traces of Fig. 2B (and described in detail Section 2.3.3)
we chose the db2/D3 wavelet and used the rectified,
downsampled coefficient trace to identify time of
occurrence of EMG bursts.
2.3.3. Choice of wavelet shape and scale
As in principal components analysis and Fourier
analysis, wavelets form a set of basis functions that
can be used for reconstruction of a signal. In contrast to
principal components analysis, the bases are generally
chosen in advance of the analysis, rather than being
derived from the data set. In contrast to Fourier
analysis, the bases are vectors of discrete length */thus
wavelet analysis is most appropriate for reconstruction
of a non-periodic signal (such as the unitary EMG
bursts considered here). In analogy to Fourier analysis,
the set of basis functions is a family of wavelets,
differing in length (or ‘scale’) and position along the
time axis of the signal. In the case of the EMG signal:
EMG(t)Sj;k cj;k wj;k (t) (2)
where c and w are the wavelet coefficient and the
wavelet, respectively, and j and k represent the scaling
and shifting of the wavelet. If the wavelet starts at time
t /0 and ends at time t/N , then the rescaled wavelets
wj 0 start at time t/0 and end at time t/N /2j . The
shifted wavelets w0k start at time k and end at time k
/
N.
Thus the EMG signal can be decomposed into or
reconstructed as a weighted sum of wavelets of all
possible lengths (‘scales’), occupying all possible posi-
tions in time. We used Matlab Wavelet Toolbox (The
Math Works, Inc.) to perform both continuous and
discrete versions of this transformation. As shown in
Fig. 3, we began by processing our EMG signals with a
continuous wavelet transform. The continuous trans-
form displays the temporal structure of the various
different frequency components of the signal. The y -axis
is the wavelet scale, with long wavelets at the top
(representing the lowest frequency components) and
short wavelets at the bottom (representing the highest
frequency components). Large positive coefficients are
shown in white; large negative coefficients are shown in
black.
We tried a number of different wavelets. The filters
available in the Toolbox range in shape from a square
wave, to more triangular waveforms, to Gaussian and
Mexican Hat shapes. Within a given family of wavelets
(e.g. the Daubechies family) the waveforms can range
from simple and primarily biphasic, to much smoother
and multiphasic forms. In Fig. 3, we present an example
where a 200 ms segment of biceps EMG was subjected
to a continuous wavelet transformation. In the upper
plot, it is apparent that the Gaussian (bell-shaped) filter
did not fit the EMG signal well, especially at the longest
wavelet scales. This is evidenced by the lack of a clear
unitary peak (either positive or negative) at the time of
the EMG burst.
The Haar is the simplest wavelet; it is a biphasic
square wave. Haar coefficients for reconstructing our
sample EMG trace are shown in the middle plot of Fig.
3. The Haar wavelet fit the largest fluctuations from
negative to positive (or vice versa), at a wide range of
scales.
The Daubechies family of wavelets seems to resemble
motor unit potentials most closely (see Fig. 2A). The
simplest of these wavelets is Daubechies 2, or db2 (db1 is
actually the Haar wavelet). Wavelet db2 is a triphasic
triangular wave, with a small negative triangular pulse
preceding a large positive triangular pulse and then a
medium-sized negative pulse (inset, Fig. 2A). The
Daubechies 4 wavelet (db4) is substantially smoother
than db2 and has more zero-crossings (not shown). In
the bottom panel of Fig. 3, we show that db4 fit the
biceps EMG signal in a slightly different way than the
 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
Fig. 3. Continuous wavelet transformation of a 200 ms biceps EMG
segment. The values of the wavelet coefficients are plotted in gray
scale, with white representing the largest positive values (i.e. strong
positive correlations between the wavelet shape and the EMG shape)
and black representing the largest negative values (where a negative
image of the wavelet gave a good fit). All three filters (Gaussian, Haar,
and Daubechies) identified the main burst with a sequence of high
negative and positive coefficients. There were, however, differences
across wavelet scale (y -axis). For example, the low frequency
components (long wavelets, A4 and D4 scales) were not well fit by
the Gaussian curve.
Haar square wave or the Gaussian curve. Large positive
and negative correlations were found at a range of
scales, with the most distinct peaks occurring at middle
to short wavelet lengths. In Fig. 4A we show the
continuous wavelet transformation for the same biceps
169
signal, using db2 instead of db4. The db2 spectrum was
intermediate to the Haar and db4, in that it clearly fit
the large voltage fluctuation at burst center and did so
over a wide range of frequencies. In the middle range of
wavelet lengths (D3) there were two distinct positive
peaks in the coefficient trace corresponding to the times
of occurrence of the two multiunit bursts.
For the purpose of the present study, we wished to
focus on only one wavelet at only one scale. We chose
the db2 wavelet for its similarity to the shape of motor
unit potentials (see Fig. 2A) and its relative simplicity.
We used this wavelet in a discrete analysis, as dia-
grammed in Fig. 4B. This progressive splitting of the
EMG signal into approximation and detail resulted in
wavelet coefficients at five scales: A4, D4, D3, D2 and
D1. The EMG signal can be perfectly reconstructed
using weighting coefficients specific to each scale (Strang
and Nguyen, 1997).
In Fig. 4C we show the peak-to-peak range of the db2
weighting coefficient values for several muscles (subject
B). The D3 component had the largest peak weighting
coefficients. The length of the db2/D3 wavelet is about
20 /30 ms, the approximate length of a motor unit
action potential (see Fig. 2A). (The approximate length
of D1 is 6 ms, D2 is 12 ms, and D4 is 48 ms.) It is
therefore reasonable that the D3 scale contained the
largest positive and negative correlations with EMG
bursts.
2.4. Application of wavelet analysis to single-trial EMG
We adopted the db2 wavelet and examined traces
representing the weighting coefficients for a middle
scale, D3. We then recorded the time of the largest
peaks in the coefficient trace, as illustrated in Fig. 5.
Trials were excluded when no peaks exceeded the
background amplitude by more than 50% (this tended
to limit our analysis to trials with movement times less
than about 700 ms). For each set of data (subject,
muscle, direction), we marked the two largest bursts.
Fig. 5 shows that the peaks of the coefficient trace were
often different from the peaks of the rectified EMG
signal.
To further aid in rejecting spurious peaks, we set
restrictions on each analysis that tended to confine the
search of each coefficient trace to one of two patterns.
This was especially important in cases (about 20% of all
trials) where background (or ‘tonic’) EMG levels
showed a large increase during the movement (see
Section 4). In an ‘agonist pattern’ one burst was marked
in the period centered 100 ms prior to movement onset
and a second burst was found in a period ending near
the time of movement end (see biceps activity in Fig. 5).
In an ‘antagonist pattern’ two bursts were marked in the
middle section of the movement time (see medial triceps
activity in Fig. 5). To avoid marking a single burst as
170 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177

Fig. 4. Rationale for choosing the Daubechies 2 (db2) wavelet at a mid-range scale (D3). In panel A, we show the continuous wavelet transform for
the same biceps EMG signal as shown in Fig. 3. Although the burst is highlighted at a wide range of wavelet scales, the data in panel C show that for
all muscles, the middle scale (D3) had weighting coefficients with the largest peak-to-peak amplitudes. These data are averages (9/standard error) for
(n/29) trials from subject B (outward movements). In panel B, we diagram the discrete wavelet transformation. The EMG signal can be
reconstructed as a sum of five components (A4, D4, D3, D2, D1).

two bursts, we imposed the restriction that successive 3. Results

peaks must be more than 50 ms apart. As we will
demonstrate below, there was no clear division between
the timing of ‘agonist’ and ‘antagonist’ bursts, and thus
this ‘antagonist activity’ was identified within a large
time window, slightly overlapping the first and second
agonist time frames (as also shown in Flanders, 1991;
Flanders et al., 1994, 1996).
3.1. Time frames
‘Agonist activity’ generally preceded the triggering of
switch #1 by about 100 ms and ‘antagonist activity’ was
generally centered around a time 100 ms prior to the
middle of the movement. These times are marked by
 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
Fig. 5. Identification of burst timing in biceps and medial triceps.
Although rectified EMG and coefficient traces are shown, the wavelet
filter was applied to unrectified EMG data. The rectified coefficient
trace was then searched for the times of the highest peaks. To narrow
the search, the period between the first and second agonist burst was
masked for muscles that fit the agonist pattern (such as biceps in this
example). For muscles (e.g. medial triceps) that did not fit this pattern,
we marked two bursts in an intermediate time frame. The vertical
dashed lines indicate the switch times, and for reference we also show
solid vertical lines 100 ms prior to movement onset and middle. All
data are from subject A, for a single outward movement.
solid vertical lines in Fig. 5. This figure gives an example
of the relation of EMG bursts to movement time, using
rectified data from a single trial where the movement
direction was ‘out’ (subject A). Biceps had bursts at the
beginning and the end of the movement; medial triceps
began to burst shortly after movement onset. Although
in this example the first triceps burst was larger than the
second, it was not unusual to find two bursts of
approximately equal amplitude in the antagonist time
frame. The traces below each EMG represent the
rectified db2 wavelet coefficients. The two peaks identi-
fied for each muscle are marked by circles.
3.2. Quantification of burst timing
The timing of EMG bursts scaled in linear relation to
movement time (the time between switch #1 and switch
#2). To quantify this relation, for each subject, muscle
and direction, we plotted the time of each of two EMG
bursts against the movement time. In Fig. 6, we
illustrate several regressions using data from subject
171
A’s medial triceps. Since the EMG data were aligned at
movement onset, it is expected that data from a burst
occurring around this time (e.g. the ‘1st Agonist Burst)
should be fit with a slope near zero. In contrast, data
from a burst occurring at the end of the movement (e.g.
the 2nd Agonist Burst’) should be fit with a slope of one,
and ‘Antagonist Activity’ should be associated with
intermediate slopes.
Because they were close to zero, regression slopes for
early EMG bursts rarely reached statistical significance.
For example, in Fig. 6 the first triceps burst for
downward movements was related to movement time
with a slope of /0.07, which, as would be expected, was
not significantly different from zero (P /0.05). How-
ever, the 95% confidence interval on the slope was small
(9/0.14), indicating that the wavelet method reliably
identified this burst. In contrast, regression slopes for
the bursts occurring after movement onset were gen-
erally significantly different from zero at the P B/0.01
level, but because variance often increased, the slopes
had a wider confidence intervals. Thus in this study, the
fits were considered ‘significant’ if the confidence
interval on the slope of the regression was less than 9/
0.25 or if the correlation coefficient was significant at
P B/0.01. Of the 396 regressions performed (11
muscles/2 bursts /6 directions /3 subjects) only 24
failed to meet at least one of these criteria. This indicates
that the wavelet approach was successful in identifying
bursts that were consistent from trial to trial, over a
wide range of speeds.
3.3. Examples of timing and scaling
Fig. 6 illustrates the timing and the time-scaling of
medial triceps for movements in 6 directions. For
movements ‘out’ (upper left panel) medial triceps burst
in the time frame of an antagonist; for movements
‘back’ (upper right panel) the pattern was that of an
agonist. The patterns for ‘up,’ ‘down’ and ‘right’ were
similar to one another, generally resembling the pattern
of an agonist. The burst times for leftward movements
were harder to classify: the first burst was in the early
antagonist time frame, but the second burst closely
followed movement end, as would be expected of the
second agonist burst. It is apparent from this example
that a given muscle cannot be easily classified as either
an agonist or an antagonist for each movement direc-
tion. This was a general conclusion for all subjects and
all muscles.
The eleven muscles did not burst in phase with one
another. This is shown in Figs. 7 and 8 (subject A). In
Fig. 7A, we begin by presenting rectified, averaged (4
trials) and smoothed (two-sided exponential filter) EMG
traces for the outward direction. The averages were
computed by aligning (at the time of switch #1) data
from the trials with movement times ranging from 324
172 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177

Fig. 6. Medial triceps EMG: burst time vs. movement time for six movement directions (subject A). Using data from single trials, the relation
between burst time and movement time was well fit by linear regression. The slopes of these lines reveal the scaling of the EMG time base. For this
subject’s medial triceps, the scaling pattern differed for different movement directions. Medial triceps also burst in different time frames for different
movement directions.

to 400 ms. As in Fig. 5, the solid vertical lines represent
the ‘expected’ time of the first agonist and antagonist
bursts. The earliest burst is that of biceps (Bic), followed
in close succession by bursts in brachialis (Brac) and
bradioradialis (Brad). The shoulder flexors peaked next:
anterior and medial deltoid (AD and MD) peaked
approximately together, followed by a double burst in
pectoralis (Pec). As is often the case with averaged
waveforms, it is not clear whether the long burst in Pec
represented two bursts, a doublet, or an artifact of
averaging data from trials with a range of movement
times (the latter was in fact the case). Interestingly, the
three heads of triceps (LoT, LaT and MT) exhibited
very different EMG waveforms.
In Fig. 7, the pattern for outward movements can be
contrasted with the pattern for downward movements.
In Fig. 7B we show 6-trial averages with movement
times ranging from 325/393 ms. Here, the pattern
begins with activity in posterior deltoid (PD), presum-
ably to counteract the shoulder flexion torque that will
be the mechanical effect of elbow extension (this activity
is in phase with a quick inactivation of Brac, an elbow
flexor). Successive peaks then occur in the other
shoulder extensor (LaD), in all single-joint elbow
muscles, and then in Pec, AD, and finally MD. In
contrast to the situation shown for the outward move-
ments in Fig. 7A, for downward movements, biceps and
the long head of triceps show a simpler, more reciprocal
pattern.
Using the analysis illustrated in Figs. 2 /6 (on
unrectified single-trial EMG), the sequencing of the
bursts illustrated in Fig. 7 was quantified as shown in
Fig. 8. Fig. 8 also shows burst sequences for the other
four directions. In this Figure we use regression lines
such as those illustrated in Fig. 6, but in this case
movement time is on the y -axis and the time of the
EMG burst is on the x -axis. The plot for ‘out’ (top left
panel) can be seen to correspond to the averages shown
in Fig. 7A. The pattern begins with a burst in the flexors
(reds/orange/brown/black, and thick lines) and has most
of the extensor activity in the middle time range (greens
and blues, and thin lines). In contrast, for ‘down’
 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177 173

Fig. 7. Rectified and smoothed EMG averages for two movement directions (subject A). Averages for the outward movements (A) are composed of
data from four trials, with movement times ranging from 324 to 400 ms (as indicated by the bar on the time scale). Averages for the downward
movements (B) are composed of data from 6 trials, with movement times ranging from 325 to 393 ms. To focus attention on the EMG waveform, all
EMG traces were scaled to equate peak-to-peak amplitude. EMG traces representing flexors are depicted as thick lines; EMG traces representing
extensors are depicted as thin lines. The top row features data from the three portions of deltoid: anterior deltoid (AD), medial deltoid (MD), and
posterior deltoid (PD). The middle row contains data from the muscles that cross both the shoulder and the elbow (biceps, Bic, and the long head of
triceps, LoT) and two other shoulder muscles (pectoralis major, Pec, and latissimus dorsi, LaD). The bottom row contains data from the four elbow
muscles: brachioradialis (Brad), brachialis (Brac), the lateral head of triceps (LaT) and medial triceps (MT). As points of reference, hypothetical times
for the first agonist burst and the antagonist burst are indicated by vertical lines. Even if variations in waveform and amounts of background (anti-
gravity) activity can be overlooked, bursts of muscle activation clearly do not follow a simple biphasic or triphasic form.

(middle right panel) the pattern begins with an early
burst in posterior deltoid (PD, thin blue line) for all
movement times (300 /600 ms). As will be discussed
below (under Section 3.5) it is apparent that muscle
activation is asynchronous for all directions.
3.4. EMG time-base scaling
We used the slopes of the regression lines to test
hypotheses concerning the scaling of the EMG time
base. Notice that the slope was approximately zero
when the burst was near movement onset (a vertical line
in the plots of Fig. 8, or a horizontal line in Fig. 6). If
EMG time base scales exactly with movement time, a
burst occurring at movement end should be fit with a
slope of 1.0 (or /458 in Fig. 8). Likewise, a burst
occurring at the middle of the movement should have a
slope of 0.5 (see bottom left panel in Fig. 8).
Based on this assumption, we calculated the deviation
from the expected slope by adjusting the observed slope
according to the time of the burst. This ‘deviation from
expected slope’ was calculated such that it would be zero
if EMG time base scaled exactly with movement time.
For example, for outward movements, the first medial
174 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177

Fig. 8. Timing and time-scaling of arm muscle activation for movements in six different directions (subject A). A hypothetical ‘triphasic’ pattern is
shown in the lower left panel, with EMG time base scaling in proportion to movement time. Although muscle activation generally shows this pattern
of time base scaling, the sequencing across muscles is not triphasic, but is instead more continuous or multiphasic. Elbow flexors are brachioradialis
(Brad), brachialis (Brac). The elbow and shoulder flexor is biceps (Bic). The shoulder flexors are pectoralis major (Pec), anterior deltoid (AD), and
(sometimes) medial deltoid (MD). Shoulder extensors are posterior deltoid (PD) and latissimus dorsi (LaD). The two-joint extensor is the long head
of triceps (LoT), and the elbow extensors are the lateral head of triceps (LaT) and medial triceps (MT).

triceps (MT) burst scaled with a slope near zero (thin, deviation from expected slope was
/0.11. However,
bright green line in the top left panel of Fig. 8, see also the second MT burst did not scale as much as expected;
Fig. 6). This is approximately as expected, given its time the deviation from expected slope was large and
of occurrence, just prior to movement onset. The negative (/0.49). In contrast, the bursts in anterior
 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
deltoid and pectoralis (AD and Pec, thick pink and red
lines) tended to scale more than expected (with slope
deviations of
/0.69 and
/0.28, respectively, for the
first burst). This contrast between the scaling of medial
triceps and the scaling of the shoulder flexors was found
in two of the three subjects. Based on the 95%
confidence intervals of the slope deviations, in subjects
A and C the second burst in MT scaled significantly less
than the first burst of AD and Pec. This result
approximately replicates the phenomenon of unequal
scaling reported previously for upward and straight-
forward movements (Buneo et al., 1994). However this
was not found, even qualitatively, in the third subject.
Using this approach we tested several hypotheses
concerning differential time-base scaling. In each case
we got negative results: deviations from the expected
slopes were not muscle specific, as might be the case if
some muscles had longer tendons or greater series elastic
compliance than others. Agonists and shoulder muscles
did not consistently scale more than antagonists and
elbow muscles. We also wondered if slope deviation was
related to the intensity of each muscle’s activity, as it
varies across movement direction. To test for this
possibility we regressed the slope deviation against
EMG amplitude. Before pooling the data, for each
subject and each muscle we normalized the peak
(smoothed and averaged) EMG amplitude (from data
processed as in Fig. 7), to the peak recorded across the
six directions. The regression showed that slope devia-
tion was not well related to EMG amplitude (r2 /0.008,
P /0.05). Thus we conclude that EMG time base scales
approximately with movement time, with no consistent
exceptions.
3.5. Burst sequencing
The regression of burst time vs. movement time also
provided an estimate of the sequencing across muscles
for each movement direction. A hypothetical pattern is
drawn at the bottom of Fig. 8. Choosing, as an example,
a movement where flexors are agonists and extensors are
antagonists, we have drawn early agonist activity in
shoulder muscles to slightly precede agonist activity in
elbow muscles, with the reverse occurring for the early
antagonist activity (Karst and Hasan, 1991). As men-
tioned above, the pattern observed in each of the six
directions was not as clearly ‘triphasic’ as this hypothe-
tical example. Only in the case of the upward movement,
did activity cluster into the first agonist time frame. This
was true for subjects A (shown here) and B, but subject
C showed an asynchronous pattern even for this
direction. For these upward movements however, both
the flexors and the extensors burst prior to movement
onset and there was no antagonist burst. (Instead there
was a phasic inactivation of the flexors at the expected
time of the antagonist burst, such that the braking force
175
was that provided by gravity (Flanders et al., 1994).)
The general phenomenon of asynchronous activation is
readily apparent in Fig. 8; this phenomenon was
previously documented using very different analytical
methods (Flanders, 1991; Flanders et al., 1996).
4. Discussion
We found the wavelet analysis to be useful in
recognizing the times of occurrence of a particular event
in the EMG signal. Wavelets have been previously
employed in several other biological studies, ranging
from pattern detection in the electrocardiogram (Ka-
dambe et al., 1999) to de-noising for the purpose of
refining the information content of neuronal spike trains
(Laubach et al., 2000). Wavelet analysis has also been
used to pick out the waveform of a SMU in single unit
recordings (Fang et al., 1999), and this technique is
similar to the template matching that we have used
previously to identify SMUs (Herrmann and Flanders,
1998).
One difference between unit waveform template
matching and wavelet analysis is that the wavelet
analysis decomposes the signal into frequency bands,
i.e. components that are best fit by long or short
versions of the same wavelet. With the db2 wavelet
and a signal sampled at 1000 Hz, the smallest scale in a
discrete wavelet analysis (D1) uses a wavelet that is
about 6 ms wide. The D2 wavelet is about 12 ms wide
and the D3 wavelet is about 24 ms wide. Although both
D2 and D3 have wavelengths that appear to be
comparable to the width of a SMU potential (recorded
with surface electrodes, see Fig. 2) we found that D3
represented the most power in the signal (Fig. 4C).
Because the wavelet analysis goes through series of steps
of downsampling, the D3 level had a temporal resolu-
tion that was inferior to the resolution at the D2 level
(see Figs. 2 and 4B). However given the variability
inherent in the timing of multiunit bursts (Fig. 6) this
loss of resolution in the weighting coefficient trace did
not appear to be a major problem.
We found that the simplest wavelets (db1-2) did a
good job of recognizing multiunit EMG bursts (Figs. 3
and 4A). A slightly different wavelet (such as db4) or a
different scale (e.g. D2) would be expected to mark a
slightly different part of the burst. However, this would
be of little consequence for the study of time-base
scaling, since the main goal was to identify the same
event each time it occurred in the EMG signal. This was
apparently successful since the regression analysis re-
vealed the recognized event to be well related to move-
ment time. Although we did not systematically compare
the two approaches, the example shown in Fig. 5
suggests that marking the peaks of the rectified EMG
trace would result in greater variability.
176 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
There are many different analytical approaches to the
study of EMG patterns; most rely on rectified EMG
averages. In cases where there are clear rectified bursts
with little background activity (as in studies where the
arm is supported against gravity) EMG bursts can be
scored for onset, end and integrated area (e.g. Gottlieb
et al., 1989). In other cases, the EMG amplitudes of a
collection of muscles (Tresch et al., 1999; Spencer and
Thelen, 1999) or the EMG waveforms of a given muscle
(Flanders, 1991; Kargo and Giszter, 2000; Giszter 2001)
may be subject to some sort of data reduction, decom-
position, or component analysis. The various ap-
proaches tend to be complementary, since each focuses
on a different aspect of the signal, and none are without
limitations.
The present study represents a rare attempt to
measure EMG data from individual trials. One limita-
tion of the present analysis is its focus on the timing of a
small number of positive bursts, rather than the entire
waveform. Smaller bursts and periods of phasic inacti-
vation were not recognized. Previous analysis has
suggested that these periods of phasic inactivation are
an important part of the pattern */phasic inactivation is
often the earliest event in the pattern that subserves a
reaching movement (Flanders et al., 1994, 1996) and a
similar phenomenon has been reported for other move-
ments as well (Hufschmidt and Hufschmidt, 1954;
Gottlieb et al., 1970). Another important period of
phasic inactivation occurs between the first and the
second agonist bursts. This generally occurs near the
time of peak velocity, when the agonist muscle is
engaged in a high-speed shortening contraction. Com-
bination of EMG waveform data with a model of
muscle mechanics reveals that, due to the force/velocity
properties of muscle, this pause between the first and
second agonist bursts often allows gravity to provide a
large braking component to balance of forces (Soechting
and Flanders, 1997).
Another limitation of the present approach is that it
lacks the ability to subtract away the background
activity that is related to supporting the arm against
the pull of gravity. Ever since this subtraction approach
was justified (by a principal component analysis;
Flanders and Herrmann, 1992) our surface EMG
studies have focused on the phasic component isolated
from rectified and smoothed EMG averages. In the
present study, we avoided the problem of background
(or anti-gravity) activity by masking the time period
between the first and second agonist bursts. However, in
waveforms with a large amount of background activity
(see, for example, deltoid activity in Fig. 7A) it was still
difficult to identify a small number of relatively large
bursts. Since the subtraction procedure requires a large
number of trials with comparable movement times for
averaging, it was not possible in the present study.
Obtaining a large number of identical trials is also
difficult in animal research. The wavelet filter provides a
viable alternative to averaging when the experimental
design dictates an analysis of single trials. It also may
provide clues about the structure of multiunit activity
and a link between surface EMG and single motor unit
studies.
Acknowledgements
I thank Drs Sara T. Murray, John F. Soechting, and
Gilbert Strang for help and support. This work was
supported by a grant from the National Institute of
Neurological Disorders and Stroke, R01 NS27484.
References
Angel RW. Electromyography during voluntary movement: the two
burst pattern. Electroenceph Clin Neurophysiol 1974;36:493 /8.
Atkeson CG, Hollerbach JM. Kinematic features of unrestrained
vertical arm movements. J Neurosci 1985;5:2318 /30.
Buneo CA, Soechting JF, Flanders M. Muscle activation patterns for
reaching: the representation of distance and time. J Neurophysiol
1994;71:1546 /58.
Fang J, Agarwal GC, Shahani BT. Decomposition of multiunit
electromyographic signals. IEEE Trans Biomed Eng
1999;46:685 /97.
Flanders M. Temporal patterns of muscle activation for arm move-
ments in three-dimensional space. J Neurosci 1991;11:2680 /93.
Flanders M, Herrmann U. Two components of muscle activation:
scaling with the speed of arm movement. J Neurophysiol
1992;67:931 /43.
Flanders M, Pellegrini JJ, Geisler SD. Basic features of phasic
activation for reaching in vertical planes. Exp Brain Res
1996;110:67 /79.
Flanders M, Pellegrini JJ, Soechting JF. Spatial/temporal character-
istics of a motor pattern for reaching. J Neurophysiol 1994;71:811 /
3.
Ghez C, Martin JH. The control of rapid limb movement in the cat.
III. Agonist-antagonist coupling. Exp Brain Res 1982;45:115 /25.
Giszter, S.F. Quantization of motor activity into primitives and time-
frequency atoms using independent component analysis and
matching pursuit algorithms, Proc. IEEE/EMBS, Istanbul, Turkey,
2001.
Gottlieb GL, Agarwal GC, Stark L. Interactions between voluntary
and postural mechanisms of the human motor system. J Neuro-
physiol 1970;71:811 /3.
Gottlieb GL, Corcos DM, Agarwal GC. Strategies for the control of
voluntary movements with one mechanical degree of freedom.
Behav Brain Sci 1989;12:189 /250.
Herrmann U, Flanders M. Directional tuning of single motor units. J
Neurosci 1998;18:8402 /16.
Hoffman DS, Strick PL. Step-tracking movements of the wrist. IV.
Muscle activity associated with movements in different directions. J
Neurophysiol 1999;81:319 /33.
Hollerbach JM, Flash T. Dynamic interactions between limb segments
during planar arm movement. Biol Cybern 1982;44:67 /77.
Hufschmidt HJ, Hufschmidt T. Antagonist inhibition as the earliest
sign of a sensory-motor reaction. Nature 1954;174:607.
Kadambe S, Murray R, Boudreaux-Bartels GF. Wavelet transform-
based QRS complex detector. IEEE Trans Biomed Eng
1999;46:838 /48.
 M. Flanders / Journal of Neuroscience Methods 116 (2002) 165 /177
Kargo WJ, Giszter SF. Rapid correction of aimed movements by
summation of force-field primitives. J Neurosci 2000;20:409 /26.
Karst GM, Hasan Z. Timing and magnitude of electromyographic
activity for two-joint arm movements in different directions. J
Neurophysiol 1991;66:1594 /604.
Laubach M, Wessberg J, Nicolelis AL. Cortical ensemble activity
increasingly predicts behaviour outcomes during learning of a
motor task. Nature 2000;405:567 /71.
Nishikawa KC, Murray ST, Flanders M. Do arm postures vary with
the speed of reaching? J Neurophysiol 1999;81:2582 /6.
Soechting JF, Flanders M. Evaluating an integrated musculoskeletal
model of the human arm. J Biomech Eng 1997;119:93 /102.
Soechting JF, Lacquaniti F. Invariant characteristics of a pointing
movement in man. J Neurosci 1981;1:710 /20.
Spencer JP, Thelen E. A multimuscle state analysis of adult motor
learning. Exp Brain Res 1999;128:505 /16.
177
Strang G, Nguyen T. Wavelets and Filter Banks. Wellesley, Massa-
chusetts: Wellesley-Cambridge Press, 1997.
Tresch MC, Saltiel P, Bizzi E. The construction of movement by the
spinal cord. Nature Neurosci 1999;2:162 /7.
Wadman WJ, Denier van der Gon JJ, Derksen RJA. Muscle activation
patterns for fast goal-directed arm movements. J Human Mov Stud
1980;6:19 /37.
Yamaguchi GT, Sawa AGU, Moran DW, Fessler MJ, Winters JM. A
survey of human musculotendon actuator parameters. In: Winters
JM, Woo SL-Y, editors. Multiple Muscle Systems: Boimechanics
and Movement Organization. New York: Springer-Verlag,
1990:717 /73.
Zajac FE. Muscle and tendon: properties, models, scaling, and
application to biomechanics and motor control. Crit Rev Biomed
Eng 1989;17:359 /411.