Extraction of Some Relevant Instants from EMG Signal

Soﬁa Ben Jebara

Research Laboratory COSIM, Higher School of Communications of Tunis,

University of Carthage, Route de Raoued 3.5 Km, Cit

e El Ghazala, Ariana, 2088, Tunisia

Keywords:

Pre-motor Onset, Motor Onset, Motor Offset, Muscle Preparation, EMG Signal.

Abstract:

In this paper, an algorithm to estimate key instants in EMG signal during pre-motrice and motrice phases

is developed. It detects automatically i) the onset of muscle activity when an explicite recommendation of

preparation is dictated (pre-motor onset), ii) the onset of effective muscle contraction (motor onset) and iii)

the instant of muscle contraction desactivation (motor offset).

The algorithm is based on statistical thresholding and counting the number of samples exceeding the

threshlold. The counting is ensured by elaborated temporal scanning in forward and backward directions.

The threshold calculus is based on statistics of the EMG signal during muscle activity and muscle rest.

The algorithm is illustrated for the superﬁcial ﬂexor muscle during a handgrip exercice and is validated using

subjective visual inspection and objective evaluation (error rate). The results revealed that relevant instants in

EMG signal are well estimated.

1 INTRODUCTION

Electromyography is the most commonly used tool

for investigating muscle function. For example, the

ElectroMyoGraph signal (EMG) reveals details of the

timing and the magnitude of muscle activation. Re-

garding timing, an EMG signal is composed of differ-

ent kind of time intervals. Fig.1 illustrates an exam-

ple of timing. The preparation duration or pre-motor

activity or foreperiod is the time interval between a

warning signal motivating mental preparation and a

Go signal for motion execution. The pre-motor time

is followed by the motor task which is the effective

muscle activity (sometimes after a short latency time).

The activity generally ends spontaneously or after a

termination signal.

Three important instants characterize the time do-

main evolution. They are: the pre-motor onset which

characterizes the onset of pre-motor activity of the

muscle during the preparation period (initiation and

planiﬁcation of the motion), the motor onset which

characterizes the muscle activity beginning (motor

program) and the offset which characterizes the end

of the activity (see Fig.1).

Traditionally, the most reliable way to detect on-

set and offset is the visual inspection and decision of

experts like physiological therapists. Although their

accuracy, they are very expensive and time consum-

ing.

Signal processing based algorithms has been intro-

duced to overcome these drawbacks. Common meth-

ods for detecting muscle activity onset and offset were

based on thresholding (Ozgunen et al., 2010), energy

operators (Li and Aruin, 2005), signals decomposi-

tion and transformation such as empirical mode de-

composition (Lee et al., 2009) and many others algo-

rithms.

Although time preparation and its effect on mus-

cle contraction has been investigated since the early

1980 (see for example (Alegria, 1980)), at our knowl-

edge, there is no automatic method to detect the pre-

motor onset. In fact, this task seems difﬁcult since

EMG signal during preparation has low level and can

be confused to background noise.

In this work, we aim developing a signal process-

ing based algorithm to detect the three most important

instants of muscle activity which are the pre-motor

onset, the motor onset and the motor offset.The al-

gorithm is inspired from (Abbink, 1999), originally

conceived to detect muscle onset which is extended

here to estimate the three mentioned instants.

The paper is organized as follows. Section 2 de-

scribes some pre-processing tasks useful to prepare

the algorithm, such as: i) signal smoothing and recti-

fying, ii) muscle activity detection to separate muscle

contraction from muscle rest, iii) analysis windows

choice centered on muscle contraction. Section 3 de-

tails the different steps of the proposed algorithm: i)

Jebara, S..

Extraction of Some Relevant Instants from EMG Signal.

In Proceedings of the 3rd International Congress on Sport Sciences Research and Technology Support (icSPORTS 2015), pages 37-40

ISBN: 978-989-758-159-5

Figure 1: Instants of interest deﬁnition.

threshold calculus based on histograms, ii) counting

of samples exceeding the threshold by adopting a par-

ticular manner to scan the analysis window, iii) de-

tection of the instants of pre-motor onset, motor onset

and offset, iv) and ﬁnally corrections and adjustments

in case of burst or noise. Section 4 presents the results

using subjective comparison and quantitative evalua-

tion based on error rates calculus. Finally some con-

clusions and perspectives are given.

2 PRELIMINARY PROCESSING

After EMG recording, some preliminary tasks are car-

ried in order to prepare data for pre-motor onset, mo-

tor onset and motor offset detection.

• The DC component and very low frequency com-

ponents due to movement and other artifacts are

eliminated thanks to a high-pass ﬁltering whose

cutoff frequency is 10 Hz. Fig.2.a shows an ex-

ample of a pre-processed EMG signal composed

of 3 contractions of the superﬁcial ﬂexor of the

forearm during a handgrip exercice. The contrac-

tions last 4.4 s and are alterned with rest intervals

of 44s.

• The EMG signal is smoothed so that the steep

spikes are cut away and the signal looks like an

envelope (see Fig. 2.b). The Root Mean Square

(RMS), which reﬂects the mean power is used:

RMS(n) =

N + 1

N/2

∑

k=−N/2

x(n + k)

, (1)

where N is the sliding windows size on which

each EMG sample x(n) is smoothed. It is chosen

equal to 512 samples for a sampling frequency of

1 kHz.

• A preliminary Muscle Activity Detector (MAD)

is estimated by comparing each amplitude to a

threshold (see Fig.2.c):



MAD(n) =

1 ifRMS(n) > τ

0 otherwise,

(2)

Figure 2: EMG signal, RMS signal and associated MAD

ﬂag.

Figure 3: Histogram of the RMS signal.

Where τ is the threshold. It is obtained using the

following approach.

• A long RMS signal composed of 15 contractions

is used. It is composed of 4.4 ∗ 15 = 66 seconds

of activity and 44 ∗ 15 = 660 seconds of inactivity.

The ratio between these two durations is ten.

The histogram of this long RMS signal is cal-

culated and illustrated in Fig. 3. It contains one

signiﬁcant peak which represents samples with

low values occurring mainly during rest interval.

The small peak situated for higher RMS values

reﬂects muscle activity.

From the histogram, the position of the maximum

is estimated and is denoted µ and the standard de-

viation is denoted σ. The threshold τ is estimated

as follows:

τ = µ + γ.σ, (3)

where γ is a regulation parameter adjusted empir-

ically. For muscle activity detection, the experi-

ment showed that it can be chosen equal to 0.5.

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

3 KEY INSTANTS DETECTION

• According to the MAD, each contraction is cen-

tered on a window composed of the half of the

rest interval preceding it and the half of the rest

interval following it. It is called the search inter-

val (see Fig.2.b).

• The work begins in the center of the search inter-

val. Traveling in the direction of the beginning of

the interval, a counter index is associated to each

position i of the selected window:

C(i) = n

(i) + n

(i), (4)

where n

(i) (resp. n

(i)) is the accumulated

number of RMS values under (resp. above) the

threshold τ from position 1 to i (resp. from

position i to the end). Experimentally and in this

part of the work, the threshold is obtained for

γ = 3 in Eq. 3.

The same process is done by coming back to the

center of the search interval and traveling in the

direction of the end of the search interval.

• Fig.4 (resp. 5) illustrates the counter index and its

ﬁrst order derivative in the case of a contraction

without (resp. with) preparation. One can notice

from Fig. 5 that the counter index increases un-

til the contraction preparation begins, it then de-

creases until the effective contraction begins. It

remains constant until the end of the contraction

and ﬁnally, it increases again.

• The decision to detect the instants of preparation

(pre-motor), onset and motor offset are based on

the derivative of the counter index (lower parts in

the previous ﬁgures). The rule is straightforward:

if the derivative changes from +1 to -1, the pre-

motor onset is detected. If the derivative changes

from -1 to 0, the motor onset is detected and ﬁ-

nally, if the derivative changes from +1 to 0, the

motor offset is detected.

• Unfortunately, some undesirable short contrac-

tions and bursts can arise during EMG acquisi-

tion (see for example Fig. 6). They generate, in

the counter index, some local deviations from the

right way. To avoid them, a smoothing, using for

example the median ﬁlter, is applied on counter

index.

• Moreover, if many onset and offset candidates ap-

pear in one analysis window, only one effective

onset and only one effective offset must be re-

tained, they are those who correspond to the maxi-

mum area in the RMS signal. In fact, the effective

contraction lasts more and have more energy than

local burst due to noise or undesirable short con-

tractions.

Figure 4: The RMS signal, the counter index and its deriva-

tive for a contraction without muscle preparation.

Figure 5: The RMS signal, the counter index and its deriva-

tive for a contraction with muscle preparation.

Figure 6: The RMS signal, the counter index and its deriva-

tive for a contraction with short contractions and bursts.

4 RESULTS

To conduct the study, a particular experimental proto-

col was deﬁned. 24 young motivated adults (12 male

and 12 female) performed a hand grip motor perfor-

mance. We chose the superﬁcial ﬂexor to validate the

method.

Two separate guidelines were presented to the par-

ticipants:

• an auto-initiated (self-made) mode: the start and the

end of the contraction are managed by the subject it-

self, without preparation set-point.

• an external triggered mode: a preparation directive

is given (”get ready”) followed by an execution direc-

tive (”Go”) after 6.6 seconds.

For each mode, 5 consecutive hand-grip contrac-

tions are carried. Each one lasts 4.4 seconds and is

followed by a long period of rest (44 seconds).

Extraction of Some Relevant Instants from EMG Signal

Table 1: Error rate of instants estimation.

Instant pre-motor onset motor onset offset

Total error rate (%) 17 3.3 2

Figure 7: Illustration of relevant instants detection.

The EMG signals are recorded using EL508 system

which is produced by the BioPac company. Signals

are sampled at 1kHz.

Fig. 7 shows an example of muscle contraction

and the detected instants. From visual inspection, one

can see that pre-motor onset, motor onset and offset

are well detected.

From the database of the 240 contractions, we cal-

culate the relative error as the difference, in absolute

sense, between the estimated instant and the one de-

termined by visual inspection. The error rate is re-

sumed in Tab. 1. One can see that error rate is very

low for motor onset and offset (3.3% and 2% respec-

tively). It is however, slightly higher for pre-motor

onset (17 %). This last result is predictable since pre-

motor phase is characterized by low level signal and

can be confused to acquisition noise and muscle rest.

When they are non null,the values of errors, in mil-

liseconds, are:

• motor onset: 36, 260, 344, 383, 1002, 1633 and

1708,

•: offset: 112, 114, 115, 119 and 167,

• pre-motor onset: the number of errors is higher,

thats why their histogram is used for illustration (Fig.

Figure 8: Histogram of pre-motor onset estimation error.

8). One can see that the error range value is quite

large (from 0 to 6500 ms). It means that when an er-

ror occurs, the estimation can be either quite precise

or quiet wrong.

5 CONCLUSION

In this paper, an automatic algorithm to detect some

important instants in EMG signal has been developed.

It is based on the entire EMG signal (including rest in-

terval) to estimate a threshold and to count, in a par-

ticular manner, the number of samples exceeding it.

The results are very satisfactory in terms of motor on-

set and offset and present slight errors for pre-motor

onset detection.

REFERENCES

K. T. Ozgunen, U. Celik and S. S. Kurdak, “Determination

of an optimal threshold value for muscle activity de-

tection in EMG analysis,” Journal of Sports Science

and Medicine N9, pp. 620–628, 2010.

X. Li and A. S. Aruin, “Muscle activity onset time detec-

tion using Teager-Kaiser energy operator,” Proc. of

the IEEE Engineering in Medicine and Biology, Sept.

2005.

J. Lee, H. Ko, S. Lee, H. Lee, Y. Yoon, “Detection technique

of muscle activation intervals for sEMG signals based

on the Empirical Mode Decomposition,” Proc. of the

IEEE Engineering in Medicine and Biology Society

EMBC, 2009.

J. Alegria, “Contr

ole strat

egique du choix d’un instant pour

se pr

eparer

a r

eagir,”. Requin, Anticipation et Com-

portement, Editions du CNRS, Paris, pp. 95–105,1980.

J. H. Abbink“Muscle response to loading during rythmic

open-close movements of the jaw,” PhD thesis,Utrecht

university- the Netherlands, Sept. 1999.

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support