EMG Onset Detection

Comparison of Different Methods for a Movement Prediction Task based on EMG

Marc Tabie

and Elsa Andrea Kirchner

1,2

AG Robotik, University of Bremen, Bremen, Germany

German Research Center for Artiﬁcial Intelligence (DFKI), Robotics Innovation Center, Bremen, Germany

Keywords:

Online EMG Onset Detection, Movement Prediction.

Abstract:

In this work a study with 8 male subjects was conducted to compare three preprocessing methods for online

capable movement prediction based on the recorded electromyogramm (EMG) signals of the right upper limb.

One of the compared methods is the widely used Teager Kaiser Energy Operator (TKEO), the other two are

a recently proposed method that is based on variance calculation of the signal and the standard deviation.

Scope of the work was to show that fast methods, which are required for online processing, have at least

the same performance as more classical approaches with higher demands on computational resources like

the TKEO. An adaptive threshold was used for onset detection after preprocessing in all compared cases.

Comparisons of preprocessing methods were done with respect to the performance in movement prediction

and earliness of onset detection. The inﬂuence of different movement speeds on the prediction time and

the performance were investigated as well. Results presented here show signiﬁcant differences between the

investigated preprocessing methods concerning the prediction time. As a further result of this study it could

be shown that different movement speeds also have a signiﬁcant effect on the prediction time.

1 INTRODUCTION

Motivated by a robotic application developed by our

group that enables movement assistance by an ex-

oskeleton, we were reviewing fast methods for move-

ment prediction based on psychophysiological data.

In previous studies we could show that the control

of an exoskeleton beneﬁts from prediction of move-

ments based on psychophysiological data. It could

be shown that the electroencephalogram (EEG) can

be used to improve the interaction between human

and an exoskeleton (Folgheraiter et al., 2012). How-

ever, due to the so called electromechanical delay be-

tween measurable changes in the electromyogramm

(EMG) and the production of force in the correspond-

ing muscle (Cavanagh and Komi, 1979) (Zhou et al.,

1995) a fast detection of EMG onset activity can also

be used to predict a physical movement. There are

several studies concerning the automatic onset detec-

tion in the EMG. It is stated that preprocessing of

EMG signals with the Teager Kaiser Energy operator

(TKEO) improves the onset detection in comparison

to simply ﬁlter the EMG data (Kaiser, 1990) (Li et al.,

2007) (Solnik et al., 2010). However, fast and simple

methods that require low computational resources are

required for embedded signal processing that allow

onboard analysis. Recently a preprocessing method

that is based on the variance (VAR) calculation of

EMG signals was proposed (Nikolic and Krarup,

2011), but to our knowledge so far no comparison of

the TKEO and variance was done. Thus the compari-

son will be one goal of this paper.

After preprocessing onset detection can be per-

formed with different threshold methods, like single

threshold (Hodges and Bui, 1996) or multi thresh-

old (Bonato et al., 1998) detection. The drawback

of those methods is that the performance strongly de-

pends on the threshold that is chosen once for the

whole dataset. Recently an adaptive threshold method

was proposed, in which the threshold is continuously

updated based on the mean and standard deviation of

a sliding window containing the recent data point and

N previous data points (Semmaoui et al., 2012). Thus,

onset detection can be adapted to possible changes in

the EMG as they may occur during extensive behav-

ior due to fatigue of the muscle. To our knowledge,

this adaptive approach was not yet combined with the

above mentioned preprocessing methods for the pur-

pose of EMG onset detection. Therefore, another goal

of this paper is to investigate the performance of the

242

Tabie M. and Andrea Kirchner E..

EMG Onset Detection - Comparison of Different Methods for a Movement Prediction Task based on EMG.

DOI: 10.5220/0004250102420247

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 242-247

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

adaptive threshold in combination with the prepro-

cessing methods.

To compare these different processing methods

we did not only choose performance in detection but

also the earliness of detection of EMG activity as a

performance measure. In this context we also investi-

gated different speeds of movements. Since different

speeds of behavior do occur under natural condition,

in which embedded systems for movement prediction

might be applied and may inﬂuence predictability.

2 MATERIALS AND METHODS

2.1 Experimental Setup

Eight healthy right-handed male subjects (age 29.9 ±

3.3 years) participated in the study. The subjects were

seated in a comfortable chair in front of a table. A

monitor and two switches, a ﬂat board and a buzzer,

were located on the table. The ﬂat board consisted of

a plastic cuboid (10 × 10 ×1.5 cm) and a plastic plate

(10 × 10 × 0.4 cm). The two parts were connected

with a rotary joint. Inside the cuboid a microswitch

and a spring were mounted. When putting a hand on

the whole device, the plate will press the microswitch.

By lifting the hand again the spring will push the plate

up for approximately 0.5 cm which causes the switch

to be released again. During the experiments both in-

put devices were used to determine the begin or the

end of a movement.

The experiments consisted of intention based

movements of the right arm. The subjects were asked

to perform 40 voluntary movements starting from the

ﬂat board to the buzzer and back. The events from the

input devices (pressing/releasing) were marked in the

recorded EMG data. Experiments with normal, slow,

and fast speeds were performed. For each movement

speed three runs were recorded.

Normal movements can be deﬁned as natural

ones. We simply ask the subjects to hit the buzzer

in their own speed. These runs were performed ﬁrst

and used to make the test persons familiar with the ex-

periment and the whole setup. They were not used for

analysis. For slow movements a minimum time of 1 s

from the ﬂat board to the buzzer had to be satisﬁed.

For fast movements the subjects were asked to move

as fast as possible.

Since maximum movement speed differs between

subjects, a preliminary investigation was done. Each

subject was asked to perform the movement to the

buzzer and back starting from the ﬂat board, for 10

times as fast as possible. From the 10 movements the

mean time needed to move from the ﬂat board to the

buzzer was calculated for each test person. Those av-

erage times plus an offset of 10 ms were used as the

maximum time for fast movements. The maximum

times varried from 120 to 275 ms. The experiments

were designed and executed using Presentation (Neu-

robehavioral Systems, Inc.).

During the experiments a green circle with a black

ﬁxation cross was shown on the monitor. A resting

time of 5 s between two movements had to be main-

tained. A wrong movement was indicated to the sub-

ject by changing the color of the circle from green

to red for 100 ms. Wrong movements were deﬁned

as moving before the resting time (5 s) was expired,

moving too fast, e.g., in case slow movements were

requested, and moving too slow, e.g., in case fast

movements were requested. Wrong movements were

not taken into account for data analysis. In order to

get the same amount of movements from each test

person, a run was ﬁnished after 40 valid movements.

To determine the physical begin of a movement a mo-

tion tracking system was used to track the position

of the right hand, which is further explained in Sec-

tion 2.3.1.

2.2 Data Acquisition

EMG was acquired at 5 kHz with a bipolar setup on

four muscles of the right arm, named M. brachio-

radilis, M. bizeps brachii, M. triceps brachii, and M.

deltoideus. For the measurement Ag/AgCl gel elec-

trodes were used. The skin of the test persons was

prepared with medical alcohol to obtain better con-

ductivity. The signals were ampliﬁed and digital-

ized by a BrainExG MR (Brain Products GmbH, Ger-

many) ampliﬁer and saved to a computer. The events

from the two switches, meaning pressing and releas-

ing, were marked in the EMG data. The motion track-

ing system consisted of three ProReﬂex1000 cam-

eras (Qualisys AB, Sweeden) and a passive infrared

marker mounted on the back of the right hand of the

subjects. The position of the marker was tracked with

a frequency of 500 Hz and stored on a computer. The

accuracy of the system was approximately 0.15 mm.

2.3 Data Processing

2.3.1 Finding Physical Movement Onsets

The motion tracking system was used to determine

the physical movement onset. For this the movement

speed of the right hand was derived from the track-

ing data. Afterwards, the EMG and the tracking data

were synchronized. As mentioned before in Subsec-

tion 2.1, the begin of a movement was marked in the

EMGOnsetDetection-ComparisonofDifferentMethodsforaMovementPredictionTaskbasedonEMG

243

EMG data using a ﬂat board. Since the device con-

tains a microswitch, it is obvious that the subject was

already moving when the switch is released. How-

ever, after synchronizing the EMG data and the move-

ment speed, the markers from the ﬂat board were

used as a starting point for ﬁnding the physical move-

ment onsets in the tracking data. Starting from those

marker positions the movement speed was analyzed

backwards. Once the movement speed was below

a threshold, the detected position was saved as the

physical movement onset. This threshold was set to

0.15 mm/sample, since this is the given accuracy of

the tracking system. This was done consecutively for

all recorded runs and contained movements. Figure 1

illustrates the procedure. The vertical dashed line in-

dicates the position of the marker from the ﬂat board

and the solid vertical line indicates the found position

of the physical movement onset.

Figure 1: Physical movement onset detection. The move-

ment speed is show. The dashed vertical line indicates the

position of the marker from the ﬂat board and the solid one

the detected physical movement onset.

In the further analyses the determined physical move-

ment onsets, found with the motion tracking system,

were used as the reference for the onset detection in

the EMG data.

2.3.2 Preprocessing of EMG Data

Three different methods Teager Kaiser Energy

operator (TKEO), variance, and standard deviation

were used to preprocess the EMG data. The goal

of preprocessing is to enhance the signal-to-noise

ratio of the signals. The TKEO is widely used as a

preprocessing method in EMG onset detection. The

algorithm calculates the energy of the signal in a very

efﬁcient way. Only three consecutive samples are

needed. The formula for the TKEO is given as:

Ψ(t) = x(t)

− (x(t − 1)x(t + 1)) (1)

where x(t) is the current EMG sample.

As one can see, one of the required samples lays in the

future. In order to use the TKEO online, the formula

is thus redeﬁned as:

(t) = x(t − 1)

− (x(t − 2)x(t)). (2)

Before applying the TKEO, the data is ﬁltered us-

ing a 20 Hz second order butterworth high-pass ﬁlter.

The ﬁltering is done to remove motion artifacts, like

movement of cables or changes in resistance at the

electrodes. Afterwards the energy of the EMG signal

is calculated using the TKEO. At the end the resulting

energy is smoothed to remove large ﬂuctuations.

The variance method for preprocessing EMG sig-

nals is based on the following equation (Nikolic and

Krarup, 2011):

v(t) =

N − 1

∑

i=−m

(t + i) −

N − 1

∑

i=−m

x(t + i)

(3)

where N = 2m + 1 is the window length.

Here again samples do lay in the future, therefore the

formula is reformulated as follows for an online use

of the variance:

v(t)

N − 1

∑

i=−2m

(t + i) −

N − 1

∑

i=−2m

x(t + i)

(4)

When using the variance for preprocessing, a moving

window of ﬁxed length is used. The variance of the

window is assigned to the last data point. This is done

consecutively for all samples of the data, resulting in

a signal with smoothed baseline noise and enhanced

amplitudes during movement phases. The sensitivity

of this method can be adjusted by varying the length

of the window. For smaller windows the method re-

acts faster to changes. The drawback is that artifacts

are also more ampliﬁed. Thus by increasing the win-

dow size the baseline noise gets smoother but the am-

plitudes during movement phases are less ampliﬁed.

The procedure for the standard deviation is the

same as for the variance. Even if the calculations

of both methods are very similar, they have different

mathematical properties.

For later analysis the total computational time for

all preprocessing methods was measured. A compar-

ison of all methods is shown in Figure 2.

2.3.3 Onset Detection

For onset detection in the preprocessed EMG data an

adaptive threshold approach was used. The formula

is given as:

T (t) = x(t)

+ pµ(t)

(5)

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

244

(a) Raw EMG signal.

(b) Preprocessing with the variance.

(d) Preprocessing with the TKEO.

Figure 2: Resulting signals after the application of prepro-

cessing methods: a) original EMG signal, b) after variance,

c) after standard deviation, and d) after TKEO.

with x the mean value, µ the standard deviation, N the

length of the window for the mean and standard devia-

tion and p the sensitivity factor of the threshold (Sem-

maoui et al., 2012). The preprocessed data was passed

to the adaptive threshold procedure using various val-

ues for the parameters of the different methods. For

the standard deviation and variance window sizes of

100, 250, and 500 samples were used. For the thresh-

old the windows of sizes 5000, 10000, 15000, and

20000 samples and p values of 0 − 19 were used.

As performance measures for preprocessing methods

balanced accuracy and prediction time were used.

Balanced Accuracy. The balanced accuracy metrics

is calculated as the mean of sensitivity and speci-

ﬁcity (Velez et al., 2007). The metrics is deﬁned

as:

Balanced accuracy

def

= 0.5 ∗ [

T P

T P+FN

T N

T N+FP

]

where TP, TN, FP and FN represent the number of

true positives, true negatives, false positives and

false negatives, respectively. A classiﬁcation was

counted as TP, if the signal exceeded the thresh-

old in range 500 − 0 ms before a physical move-

ment, classiﬁcations in other ranges were counted

as FP. Not detecting a movement was counted as

FN. The number of TN was counted as the amount

of periods between two movements where no FP’s

occurred. For the onset detection, parts of the data

containing wrong movements were not taken into

account. Therefore the whole wrong movement

plus ranges of 1 s before and after it were rejected

from the signal.

Prediction Time. The prediction time is deﬁned as

the elapsed time between the detected onset in the

EMG signal and the physical movement onsets

detected by the motion tracking system.

For all combinations of the above mentioned parame-

ters the balanced accuracy and prediction time for all

recorded runs were calculated. Afterwards, a cross

validation over all subjects was done. Thus, the mean

balanced accuracy and prediction time for seven out

of the eight subjects were calculated. From those

results the best combination of parameters, meaning

those which produce the highest balanced accuracy

with earliest prediction time, was chosen and tested

on the data from the remaining subject. This was done

separately for all preprocessing methods and subjects.

The mean trainings results for different EMG chan-

nels are shown in Table 2. The training showed that

best results could be obtained when using signals

from the M. bizeps brachii. Results were statistically

analyzed by repeated measures ANOVA with method

(VAR, STD, and TKEO) and speed (slow and fast) as

within-subjects factor.

3 RESULTS

Results of all described analysis are summarized in

Table 1. Statistical analysis showed that there are sig-

niﬁcant differences between the three compared pre-

processing methods for the prediction time, but not

for the balanced accuracy. STD has signiﬁcantly ear-

lier prediction times compared to the two other meth-

EMGOnsetDetection-ComparisonofDifferentMethodsforaMovementPredictionTaskbasedonEMG

245

Table 1: Results of all preprocessing methods on the dataset from 8 subjects with slow and fast movement speeds. PT, and

BA indicate prediction time in ms and balanced accuracy respectively.

VAR STD TKEO

slow fast slow fast slow fast

SUB PT BA PT BA PT BA PT BA PT BA PT BA

1 202 0.86 52 0.85 216 0.88 65 0.84 195 0.88 50 0.81

214 0.88 54 0.88 230 0.88 66 0.88 196 0.86 50 0.9

209 0.88 50 0.96 222 0.9 64 0.94 191 0.89 49 0.93

2 194 0.81 33 0.91 215 0.84 39 0.91 201 0.8 36 0.9

221 0.92 50 0.78 226 0.9 50 0.63 207 0.95 57 0.65

175 0.74 44 0.9 182 0.74 57 0.9 143 0.81 42 0.81

3 184 0.8 91 0.58 190 0.8 101 0.56 176 0.8 86 0.63

175 0.88 63 0.63 182 0.88 78 0.63 167 0.83 55 0.63

177 0.78 54 0.51 178 0.73 76 0.49 169 0.81 29 0.41

4 129 0.4 46 0.86 115 0.41 47 0.81 119 0.4 38 0.79

93 0.4 31 0.66 102 0.44 31 0.69 110 0.44 31 0.71

106 0.46 33 0.66 135 0.45 47 0.64 99 0.44 37 0.6

5 198 0.95 43 1.0 245 0.95 47 1.0 184 0.9 41 0.96

216 0.9 56 1.0 244 0.93 65 1.0 197 0.88 40 0.94

186 0.9 46 0.98 218 0.9 50 0.98 189 0.91 46 0.96

6 173 0.89 49 0.91 219 0.93 70 0.91 163 0.93 46 0.88

149 0.86 46 0.91 177 0.84 53 0.91 139 0.91 42 0.93

182 0.85 38 0.83 213 0.85 73 0.84 148 0.89 28 0.88

7 73 0.86 44 0.96 81 0.9 45 0.96 84 0.73 44 0.94

75 0.73 42 0.96 86 0.63 43 0.89 70 0.58 41 0.93

47 0.44 26 0.84 59 0.48 29 0.81 47 0.45 27 0.86

8 218 0.76 57 0.96 238 0.76 76 0.96 207 0.8 56 0.96

234 0.79 49 0.88 252 0.79 66 0.85 203 0.8 49 0.83

282 0.88 52 0.98 296 0.86 58 0.98 248 0.94 46 0.98

Table 2: Mean balanced accuracy for training with differ-

ent EMG channels, with all, EMG1, EMG2, EMG3, and

EMG4 representing mean of all channels, M. brachioradi-

alis, M. bizeps brachii, M triceps brachii, and M. deltoideus

respectively.

VAR STD TKEO

all 0.79 0.78 0.75

EMG1 0.66 0.65 0.62

EMG2 0.81 0.81 0.8

EMG3 0.63 0.62 0.6

EMG4 0.68 0.69 0.66

ods. [prediction time: F(2, 46) = 46.4, p < 0.001,

pairwise comparisons: VAR vs. STD = 0.001, VAR

vs. TKEO = n.s., STD vs. TKEO = 0.001; balanced

accuracy: F(2, 46) = 2.59, p = n.s., all each pairwise

comparisons: p = n.s.]

The same pattern was found for different move-

ment speeds. Signiﬁcant differences for the predic-

tion time, but not for the accuracy were found. Slow

movements lead to an earlier prediction time. [pre-

diction time: F(1, 23) = 142.1, p < 0.001, all each

pairwise comparisons: p < 0.001; balanced accuracy:

F(1, 23) = 1.57, p = n.s., all each pairwise compar-

isons: p = n.s.].

It was found that the calculation time for the

TKEO needs approximately 1.5 times the time the

other two methods need.

The following parameters were determined for the

preprocessing methods in the training phase. VAR

windows size (WS) of 100 samples with p = 10 and

20000 WS for the adaptive threshold, STD WS of

100 samples with p = 7 (subjects 1,3,5,6,8) and p =

8 (subjects 2,4,7) and 20000 WS for the adaptive

threshold, and for the TKEO p = 7 and 20000 WS

for the adapt. threshold.

4 DISCUSSION AND

CONCLUSIONS

Since the calculation time for the VAR and STD is

1.5 times faster compared to the time needed by the

TKEO, they suite better the needs for online process-

ing of EMG data with embedded systems. Although,

the STD prediction time in average is 13 ms earlier

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

246

compared to the prediction time of VAR, one may pre-

fer VAR for embedded systems, since the calculation

of the square root for the STD is computational very

expensive on such devices. However, the decision de-

pends on the application, if the earliness is very im-

portant, one would use STD for preprocessing. Fur-

ther, by combining VAR or STD with the adaptive

threshold very good results for movement prediction

could be achieved. In summary, a simple and compu-

tational very efﬁcient way of predicting movements

using EMG data can thus be realized.

For some subjects rather large variations in predic-

tion times, especially for slow movements, could be

observed. Mainly two reasons could have led to these

results. First, it is possible that the subjects somehow

pretensioned there muscles, even if they were told to

move right away without any preparation. This could

have led to an earlier movement prediction. Hence,

if the subjects did so for some movements and for

other not, this could explain the variation in predic-

tion times. Second, the only constraint for slow move-

ments was a minimum time of 1 s for the movement

from the ﬂat board to the buzzer Section 2.1. The

subjects were asked to perform the movements with a

steady speed. However, the subjects may have varied

the initial movement speed, e.g., fast start followed

by a slower movement. Movements with fast initial

speed may be detected later compared to those with

slow initial speed. Thus, the differences in predic-

tion time could be explained by the variation of initial

movement speeds.

Our results show that it is possible to predict

both slow as well as fast movements. We found

that for slow movements earlier prediction times were

achieved. Whether this is a real effect, or might be

due to experimental setup, i.e., datasets from slow and

fast movements were merged for training, cannot ﬁ-

nally answered here. Due to our focus on a real ap-

plication, parameters were not optimized for a certain

speed of movement. This was done since in applica-

tions one cannot relay on a certain movement speed

and the methods will have to deal with both, fast and

slow movements.

ACKNOWLEDGEMENTS

Work was funded by the German Ministry of Eco-

nomics and Technology (grant no. 50 RA 1011 and

grant no. 50 RA 1012). We want to thank Su Kyoung

Kim for her help with the statistics.

REFERENCES

Bonato, P., D’Alessio, T., and Knaﬂitz, M. (1998). A sta-

tistical method for the measurement of muscle activa-

tion intervals from surface myoelectric signal during

gait. IEEE Transactions on Biomedical Engineering,

45(3):287–299.

Cavanagh, P. R. and Komi, P. V. (1979). Electromechanical

delay in human skeletal muscle under concentric and

eccentric contractions. European Journal of Applied

Physiology and Occupational Physiology, 42(3):159–

163.

Folgheraiter, M., Jordan, M., Straube, S., Seeland, A., Kim,

S.-K., and Kirchner, E. A. (2012). Measuring the im-

provement of the interaction comfort of a wearable ex-

oskeleton. International Journal of Social Robotics,

4(3):285–302.

Hodges, P. W. and Bui, B. H. (1996). A comparison

of computer-based methods for the determination of

onset of muscle contraction using electromyography.

Electroencephalography and Clinical Neurophysiol-

ogy, 101(6):511–519.

Kaiser, J. F. (1990). On a simple algorithm to calculate the

‘energy’ of a signal. In Proc. IEEE International Con-

ference on Acoustics, Speech and Signal Processing

ICASSP-90. Conf, pages 381–384.

Li, X., Zhou, P., and Aruin, A. S. (2007). Teager-kaiser en-

ergy operation of surface EMG improves muscle ac-

tivity onset detection. Annals of Biomedical Engineer-

ing, 35(9):1532–1538.

Nikolic, M. and Krarup, C. (2011). EMGTools, an adaptive

and versatile tool for detailed EMG analysis. Biomed-

ical Engineering, IEEE Transactions on, 58(10):2707

–2718.

Semmaoui, H., Drolet, J., Lakhssassi, A., and Sawan, M.

(2012). Setting adaptive spike detection threshold for

smoothed TEO based on robust statistics theory. IEEE

Transactions on Biomedical Engineering, 59(2):474 –

482.

Solnik, S., Rider, P., Steinweg, K., DeVita, P., and Horto-

bgyi, T. (2010). Teager-kaiser energy operator signal

conditioning improves EMG onset detection. Euro-

pean Journal of Applied Physiology, 110(3):489–498.

Velez, D. R., White, B. C., Motsinger, A. A., Bush, W. S.,

Ritchie, M. D., Williams, S. M., and Moore, J. H.

(2007). A balanced accuracy function for epista-

sis modeling in imbalanced datasets using multifac-

tor dimensionality reduction. Genetic Epidemiology,

31(4):306–315.

Zhou, S., Lawson, D. L., Morrison, W. E., and Fairweather,

I. (1995). Electromechanical delay in isometric mus-

cle contractions evoked by voluntary, reﬂex and elec-

trical stimulation. European Journal of Applied Phys-

iology and Occupational Physiology, 70(2):138–145.

EMGOnsetDetection-ComparisonofDifferentMethodsforaMovementPredictionTaskbasedonEMG

247