CGA FAQ: Quantification of EMG

CGA FAQ: Determining the Onset of EMG

from BIOMCH-L 10-12-2001

I've got to decide how to determine the onset of EMG activity from
several lower leg muscles during different tilt platform tests. For the
dynamic tests (jumping and landing) I used a threshold determined as
percentage from the maximal amplitude from the first recorded trial to
asses the onset. Since a few miliseconds in preactivation time are not
that relevant, this seemed to be an acceptable choice. Anyway the
preactivation times weren't very repeatable. This is different for the
latency times were differences are in the range of miliseconds.
In the literature latency times are often studied, but its determination
is seldom described and many times done by visual control. I would
greatly appreciate any ideas about the methodology to use. I think I'll
test some of the proposed (if any) methods and see what happens. Maybe
as suggested by D.Rosenbaum it is very difficult to rely on a pure
automatic determination without visual control and the "mixed" way is
the one to go.

--------------------------------

First I'll cite the recommended literature:

Following paper was proposed by:

* JIM LUK (JIMLUK@CUHK.EDU.HK) who was kind enough to send it to me
(thanks a lot).
and
* NILS HAKANSSON (nahakansson@ucdavis.edu)

PAUL W. HODGES, BANG H. BUI "A COMPARISON OF COMPUTER-BASED METHODS FOR
THE DETERMINATION OF ONSET OF MUSCLE CONTRACTION USING ELECTROMYOGRAPHY"
ELECTROENCEPHALOGRAPHY AND CLINICAL NEUROPHYSIOLOGY 101 (1996) 511-519

In this paper 27 different EMG onset determination methods are studied
and compared with visual determination.

* ARTURO FORNER (A.Forner@WB.UTWENTE.NL) gave following hints.

STAUDE&WOLF (1999). "OBJECTIVE MOTOR RESPONSE ONSET DETECTION IN SURFACE
MYOELECTRIC SIGNALS".
MICERA, SABATINI & DARIO (1998). "AN ALGORITHM FOR DETECTING THE ONSET
OF MUSCLE CONTRACTON BY EMG SIGNAL PROCESSING".
Both articles were published in Med. Eng. Phys.

* JAMES GORDON (jamesgor@usc.edu)

GHEZ AND GORDON, EXP. BRAIN RES. (1987) 67:225-240
----------------------------------------

In general, it seems as if the most supported method is to use an
algorithm which determines the standard deviation of the emg signal
during "pre-tilt" and then determines onset as the time at which a
threshold of x standard deviations is reached. Most proposals point out
that anyway it would be advisable to visual check the signals and that
the results will be depending on the quality of the signal.

-----------------------------------------

Following I summarised some of the responses to my query (original
comments are between " ").

* DR. DIETER ROSENBAUM (diro@uni-muenster.de)

They followed different strategies: First they took the maximal
amplitude of the signal and followed back until reaching a defined
percentage of it. The NORAXON software offers an option at which you can
define how many standard deviations above the baseline the system will
identify the ONSET.

* DR. WARREN DARLING (warren-darling@uiowa.edu)

"Use an automatic system based on pre-tilt EMG levels (e.g., 2 or 3 s.d.
above mean pre-tilt EMG) and then visually check each trial (of course,
carefully check all trials in which there are unreasonable short or long
latencies)."

* DR. TH. JÖLLENBECK (joelle@uni-wuppertal.de)

Is the strongest defender of the "visual determination".

After long time dedicated to the EMD he states:
"My results clearly show that there is no satisfactory reliable method
with exception of visual determination."

Different publications by Dr. Jöllenbeck may be consulted inelectronic
journals below:
> ITES sportmotorik2001 -> http://www.uni-saarland.de/ites

* JAMES GORDON (jamesgor@usc.edu)

"Although I agree with Rosenbaum that human judgment is superior to a
computer algorithm, I think that the human judgment should be used to
refine and correct an algorithm. Otherwise, the criteria are ambiguous
and difficult to reproduce.
So what should the algorithm be based on. Although it may at times yield
acceptable results, percentage of a max amplitude is arbitrary and
atheoretical. It is also subject to bias (i.e., the larger amplitude the
criterion burst, the later the onset is marked). A certain level of
variability in an algorithm is acceptable, but bias should be avoided.

A better approach, I believe is to compute a standard deviation of a
baseline or steady state period, and then mark the onset when the level
rises 2 or 3 times the SD. This is reasonable since EMG is inherently
stochastic. It will produce some false positives, but they can be
checked by visual control. This does require having collected
steady-state background activity at the beginning of each trial or set
of trials.

To see an implementation of this, see Ghez and Gordon, Exp. Brain Res.
(1987) 67:225-240."

* AT HOF (a.l.hof@med.rug.nl)

Proposed following:

"... try to filter the rectified, but not yet smoothed EMG with a median
filter. That is: take sample N and m sampes before and after, sort, and
take the middle value. In Malab this is the function medfilt1. Such a
nolinear filter smoothes the 'grass' in the EMG but leaves the flanges
neatly intact. Experiment with the filter width 2m+1."

Anyway Hof stated that this method is largely dependent upon the quality
of our signals.

* GILBERT GARETH (G.P.Gilbert@livjm.ac.uk)

"I have taken a mean and standard deviation for a high pass filtered EMG
signal taken over 50-100ms prior to movement. I have then taken a
rolling mean over a maximum of 50mS and when this mean exceeds 1
standard deviation then I have taken onset as being at the first
millisecond of that period."

Again this proposal was accompanied with the advise that much depends on
how clean the signal is.
Different times could be taken in order to get better results, depending
on the signal.

* JOHN BARDEN (John.Barden@uregina.ca)

his proposal is similar to that of Gilbert Gareth

"After band-pass filtering the signal I calculate the mean of the
baseline. I don't rectify the signal so the positives and negatives are
retained. This results in a mean of zero, regardless of how much noise I
may have in my baseline. Then for each data point (I sampled at 2500 Hz)
I calculate the standard deviation from the mean value (virtually zero
when there is no activation). I then calculate the average standard
deviation value across a 50 ms window (125samples). This "moving
average" is calculated every 50 ms. You can write a program to check
when the average value for each of these windows exceeds a certain
threshold value (for e.g., 3 SD). I found the choice of threshold
somewhat arbitrary in terms of selection as it will depend on the signal
to noise ratio, and in my case looking at shoulder muscles, the signal
amplitude between muscles is quite variable (hence, you often need to
choose different thresholds for different muscles because of signal
amplitude variability and not noise). I hope this is of some help. My
understanding is that this is a fairly standard/common method for onset
determination."

-------------------------------------

I would also like to thank to other responders whose contributions cover
with those described above and are not listed for economy reasons.

* JAMES WICKHAM (j.wickham@latrobe.edu.au)

* EYAL WEISSBLUETH (eyalw@macam.ac.il)

* TED CLANCY (ted@ece.wpi.edu)
He made an interesting and rather philosophical contribution about ON
and OFF of muscle activity.

Back to FAQ