Swie Yuniarto W

01-12-2006, 04:56 PM

Dear All,

As I received severals mail asking to post responses on my

original posting, below is the response I got.

Hope will be useful.

Regards

Yuniarto Swie/ UEC -JP

---------Original Posting---------------

Dear Experts,

I am doing Chaotic analysis to EMG by taking the

Correlation Dimension, and found in most case that the

Corr. Dimension increased as the increasing of the muscle

contraction.

However in some case, it was observed that even the

contraction supposed to be weak (at least showed from the

EMG power), the Corr. Dim is relatively high (higher than

the relax stage). As noise has been removed, We consider

that there is underlying mechanism cause the high

Corr.Dimension in the relatively weak contraction.

Can anyone advise on such phenomenon and to share idea

onwhat happening behind.

RESPONSES : 3 postings (excluded mails requested to post

the responses in Biomch-L)

================================================== ==============

\#1. Dear Yuniarto:

If you are analyzing raw EMG signals, then I would expect

these signals to be high dimensional. The signal itself

is an interference pattern generated by the interference

of the electrical output of a large number of individual

motor units (which may be from a few to a few hundred,

depending on the muscle you're looking at and the level of

activation). If one can think of each motor unit as a

"degree of freedom" in the underlying system, then the

true dimension could be quite high. This could cause you

problems, as the correlation dimension algorithm is not

very good at assessing such high dimensional systems

(i.e., d > about 4 or 5). You should be looking at the

entire correlation integral (not just the slope of the

curve) to see if you really are obtaining results that

will allow you to compute a valid correlation integral to

begin with. You should also see if/how these curves

change as you vary the embedding dimension. If they do

change substantially, this could be problematic.

Also, I don't know how you are filtering your data, but

standard frequency domain linear filters (e.g. Butterworth

low-pass filter) can substantially alter the nonlinear

properties (including correlation dimension) of inherently

nonlinear signals.

There's an excellent text that deals with these issues,

recently published in it's second edition, by Holger Kantz

& Thomas Schreiber. The book is called "Nonlinear Time

Series Analysis" and is published by Cambridge University

Press (2nd Ed., 2004). These authors also offer their

analysis software (the "TISEAN" package) via a

well-documented web site:

http://www.mpipks-dresden.mpg.de/~tisean/

I hope this is helpful.

Good luck!

Jon Dingwell

/#2. F Borg/* wrote:

>

> Pax!

>

> I am afraid there are not too many experts on using

nonlinear time series analysis methods to EMG. I did

general survey of methdos some years back (available as

the first paper here -->

> www.netti.fi/~borgbros/bisoni ) and turned up only a

few devoted to EMG.

> Recurrence plot analysis seems to be one of the

favoured methods.However, if someone finds something like

correlation dimension to work that would be interesting. I

would expect fatigue would be the

natural phenomenon to investigate. In order to check your

results youwould have to map the correlation function C(x)

vs the scale parameter x (I guess

> you have used the Grassberger-Procaccia algorithm).

It is true that noise can drive up the *dimension*.

Denoising can also create artifacts.

> Only a closer view of your stuff would make it

possible to tell.

>

> Best regards

/# 3. Pax!

At low signal levels the system may play tricks. In a

recent mesurement

on the leg during quiet standing we found eg a perfect

match between

gastroc and tibialilis sEMG -- TA had a very weak signal

but it was a

sort of crosstalk from GA and when amplified and filtered

it matched

perfectly the GA sEMG!

One may also wonder what SVD filtering does to a weak

signal. (Does the

algorithm use relative thresholding and scaling?) Try to

leave out SVD

and use only LP and HP. I spotted a paper (Hu et al.,

Classification of

surface EMG signal with fractal dimension, Journal of

Zhejiang

University 2005 6B(8): 844-848, www.zju.edu.cn/jzus )

where they found

that using just LP 350 Hz seemed to produce intelligible

results for D_c

calculations (D_c found to be around 1 - 3 -- forearm

supination/pronation) for the purpose of classification.

The connection between fatigue and synchronization makes

sense.

Recurrence quantification analysis (RQA) seems eg to show

an increase in

the so called %DET parameter for increasing fatigue (there

is a short

discusion in the book Merletti & Parker, Electromyography,

IEEE Press

2004, ch. 6). In order to study possible relations between

spectrum and

RQA maps and parameters I think one would have to run

simulations woth

various types of signals and then compare the results

obtained by the

different methods of analysis. As yet I have not come

accross (as far as

I remember) such a study although it would be quite

straightforward to

do (but requires some work).

Regards Frank

================================================== ======================

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

Yahoo! Mail - supported by 10million people

http://pr.mail.yahoo.co.jp/mail_pr/

As I received severals mail asking to post responses on my

original posting, below is the response I got.

Hope will be useful.

Regards

Yuniarto Swie/ UEC -JP

---------Original Posting---------------

Dear Experts,

I am doing Chaotic analysis to EMG by taking the

Correlation Dimension, and found in most case that the

Corr. Dimension increased as the increasing of the muscle

contraction.

However in some case, it was observed that even the

contraction supposed to be weak (at least showed from the

EMG power), the Corr. Dim is relatively high (higher than

the relax stage). As noise has been removed, We consider

that there is underlying mechanism cause the high

Corr.Dimension in the relatively weak contraction.

Can anyone advise on such phenomenon and to share idea

onwhat happening behind.

RESPONSES : 3 postings (excluded mails requested to post

the responses in Biomch-L)

================================================== ==============

\#1. Dear Yuniarto:

If you are analyzing raw EMG signals, then I would expect

these signals to be high dimensional. The signal itself

is an interference pattern generated by the interference

of the electrical output of a large number of individual

motor units (which may be from a few to a few hundred,

depending on the muscle you're looking at and the level of

activation). If one can think of each motor unit as a

"degree of freedom" in the underlying system, then the

true dimension could be quite high. This could cause you

problems, as the correlation dimension algorithm is not

very good at assessing such high dimensional systems

(i.e., d > about 4 or 5). You should be looking at the

entire correlation integral (not just the slope of the

curve) to see if you really are obtaining results that

will allow you to compute a valid correlation integral to

begin with. You should also see if/how these curves

change as you vary the embedding dimension. If they do

change substantially, this could be problematic.

Also, I don't know how you are filtering your data, but

standard frequency domain linear filters (e.g. Butterworth

low-pass filter) can substantially alter the nonlinear

properties (including correlation dimension) of inherently

nonlinear signals.

There's an excellent text that deals with these issues,

recently published in it's second edition, by Holger Kantz

& Thomas Schreiber. The book is called "Nonlinear Time

Series Analysis" and is published by Cambridge University

Press (2nd Ed., 2004). These authors also offer their

analysis software (the "TISEAN" package) via a

well-documented web site:

http://www.mpipks-dresden.mpg.de/~tisean/

I hope this is helpful.

Good luck!

Jon Dingwell

/#2. F Borg/* wrote:

>

> Pax!

>

> I am afraid there are not too many experts on using

nonlinear time series analysis methods to EMG. I did

general survey of methdos some years back (available as

the first paper here -->

> www.netti.fi/~borgbros/bisoni ) and turned up only a

few devoted to EMG.

> Recurrence plot analysis seems to be one of the

favoured methods.However, if someone finds something like

correlation dimension to work that would be interesting. I

would expect fatigue would be the

natural phenomenon to investigate. In order to check your

results youwould have to map the correlation function C(x)

vs the scale parameter x (I guess

> you have used the Grassberger-Procaccia algorithm).

It is true that noise can drive up the *dimension*.

Denoising can also create artifacts.

> Only a closer view of your stuff would make it

possible to tell.

>

> Best regards

/# 3. Pax!

At low signal levels the system may play tricks. In a

recent mesurement

on the leg during quiet standing we found eg a perfect

match between

gastroc and tibialilis sEMG -- TA had a very weak signal

but it was a

sort of crosstalk from GA and when amplified and filtered

it matched

perfectly the GA sEMG!

One may also wonder what SVD filtering does to a weak

signal. (Does the

algorithm use relative thresholding and scaling?) Try to

leave out SVD

and use only LP and HP. I spotted a paper (Hu et al.,

Classification of

surface EMG signal with fractal dimension, Journal of

Zhejiang

University 2005 6B(8): 844-848, www.zju.edu.cn/jzus )

where they found

that using just LP 350 Hz seemed to produce intelligible

results for D_c

calculations (D_c found to be around 1 - 3 -- forearm

supination/pronation) for the purpose of classification.

The connection between fatigue and synchronization makes

sense.

Recurrence quantification analysis (RQA) seems eg to show

an increase in

the so called %DET parameter for increasing fatigue (there

is a short

discusion in the book Merletti & Parker, Electromyography,

IEEE Press

2004, ch. 6). In order to study possible relations between

spectrum and

RQA maps and parameters I think one would have to run

simulations woth

various types of signals and then compare the results

obtained by the

different methods of analysis. As yet I have not come

accross (as far as

I remember) such a study although it would be quite

straightforward to

do (but requires some work).

Regards Frank

================================================== ======================

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

Yahoo! Mail - supported by 10million people

http://pr.mail.yahoo.co.jp/mail_pr/