Hi Mohammad,

if you have a concrete idea which features could be representative for the movements but you just do not know in which channels, you could extract such features and afterwards try to apply feature selection. If you have a look for feature selection methdos, a wrapper approach could be suitable. However, you would have take care not to assess "single features" but the combination of all features from each channel to evaluate a channel's suitability.

Greetings, Sebastian 

Thank you Mr. Sebastian for your reply and time.

Indeed, I used the features for the investigation:  RMS for the EMG and the mean for the joint angles within the processing window. 

I meant by noisy signals those which don't contribute in the movements of interest. I have multi sEMG channels and multi hand joint angles recorded during different wrist and grasping motions. The model to be designed to find the relation between sEMG and angles is sensitive to the noisy, or non informative channels by this meaning. So, I have to include only the signals who have contributions in the movements of interest (18 of wrist and grasping movements).

The problem is the sEMG were taken from: 8 uniformly distributed channels around forearm, 2 from muscle flexor and muscle extensor Digitorum Superficialis, and 2 from Biceps and Triceps. Hence, the majority of the signals (8) don't target specific muscles.

From the visual inspection, as shown in the attached file below where RMS: root mean squared, Ch: EMG channel, and DOF: degree of freedom (joint angle), some signals are good for certain movements and the other signals are bad for the same movements. I'm trying to solve the problem by investigate the correlation between those signals. The visual inspection is tedious due to the large number of signals, movements, DOFs, and volunteers.

Mr. Sebastian :

Do you agree to follow the approach of calculating the correlation between RMS (or other feature/feature set) of 2 channels of all possible combinations of channels, for all movements, for all volunteers, averaging them, then excluding those which have low correlation with respect to the average?

Another problem: each movement of interest was repeated 6 times with a rest gesture between each consecutive repetitions. But, the rest period lasted about twice of the movement itself! This makes problems to the model under construction.

Thank you in advance.

Regards,

Mohammad Alothman

Hi Mohammad,

after knowing your experiment now more in detail i would try to evaluate the suitability of each channel by checking the reproducibility of RMS signal during the movement that was carry out. To that end you could segment each channel in 6 epochs, each of them once reflecting the movement. A good channel should have a high correlation between all epochs.

Good channels also should show a close relation to the DOF siganl. however, in that case you should be careful which measure you use to quantify the relation. E.g. Pearson's correlation only assesses linear relations which may not exist in your case.

Greetings, Sebastian

Again, thank you Mr. Sebastian for these suggestions.

I'll consider other measure like Kendall's tau and/or  Spearman coefficients.

Regards,

Mohammad Alothman

Dear Mohammad,

I am understanding from your introduction, you did a test for many sEMG muscles of the upper arm (12 muscles) and several movements of the arm. So in my opinion, firstly you should determine your purpose. And you can select the agonist and antagonist muscles of each movement according to joint action, you can find it in many English references and another one in Arabic (https://www.researchgate.net/publication/264898071_Fundamentals_of_EMG_measurement_Part_1_almbady_alasasyt_lqyas_alnshat_alkhrby_lldlat_-_aljz_alawl). 

 When you finished of this step you will start of processing the signals using low and highpass filter to exclude the signal noise, rectify, and RMS. and also normalize your EMG data. See references:

English:

Halaki, M., & Gi, K. (2012). Normalization of EMG Signals: To Normalize or Not to Normalize and What to Normalize to? doi: 10.5772/49957

Sinclair, J., Taylor, P. J., Hebron, J., Brooks, D., Hurst, H. T., & Atkins, S. (2015). The Reliability of Electromyographic Normalization Methods for Cycling Analyses. J Hum Kinet, 46, 19-27. doi: 10.1515/hukin-2015-0030

Arabic:

https://www.researchgate.net/publication/264898071_Fundamentals_of_EMG_measurement_Part_1_almbady_alasasyt_lqyas_alnshat_alkhrby_lldlat_-_aljz_alawl

Finally, if you need more accurate to detect the onset and offset time I prefer to use Teager-Kaiser energy, more information in the next papers:

Solnik, S., DeVita, P., Rider, P., Long, B., & Hortobágyi, T. (2008). Teager-Kaiser operator improves the accuracy of EMG onset detection independent of signal-to-noise ratio. Acta of Bioengineering and Biomechanics, 10(2), 65–68. http://doi.org/101194794 [pii]

Solnik, S., Rider, P., Steinweg, K., Devita, P., & Hortobágyi, T. (2010). Teager-Kaiser energy operator signal conditioning improves EMG onset detection. European Journal of Applied Physiology, 110(3), 489–498. http://doi.org/10.1007/s00421-010-1521-8

I hope my answer help you.

Best regards,

Abdel-Rahman Akl

Book Fundamentals of EMG measurement (Part 1) المبادئ الأساسية لق...

Dear Dr. Abdel-Rahman Akl,

Thanks for your reply and references. I should study the Teager-Kaiser energy method.

In fact, I already had performed the necessary processing techniques.

As mentioned in my 2nd comment in this post, the problem is the sEMG were taken from: 8 uniformly distributed channels around forearm, 2 from muscle flexor and muscle extensor Digitorum Superficialis, and 2 from Biceps and Triceps. Hence, the majority of the signals, 8 don't target specific muscles.

I'm trying to solve problem by examining the correlation between EMG signals themselves and between EMG and kinematics (joint angles).

Regards,

Mohammad Alothman