Hi Daniel,

Yes, you can extract the same features from EMG/EEG signals.

On the other hand,

The first difference/divergence point is related to how you prepare your EMG/EEG signals for feature extraction. As you know, EEG is decomposed into several bands, these are: delta (30Hz) (each with a certain frequency range that varies between publications). You can extract any time-domain, frequency-domain, or time-frequency domains features either from the original EEG signal (with full spectrum, say 0--58Hz or whatever range), or from the individual spectral bands (after filtering to get that band of interest). In comparison to EMG, we usually extract the features from one band (10Hz to nearly 500Hz). Or, you can also partition the EMG spectrum into ranges and extract features from each range just like EEG (you will need to determine the EMG ranges for your task accordingly).

The other divergence parameters could be related to how you further clean the signals, with EEG you need to clean eye-blinks and/or jaw movements away and with EMG you need to clean ECG beats (depending on where you collected the signals from). There could be other factors to worry about (depending on your application).

Other divergence points between the two signals' feature extraction process could also be related to the windows size/other parameters. For example, with EMG we usually use 100 msec to 150 msec windows when working on prosthesis control and may be larger values for different applications (usually small values depending on the application). On the EEG side, this will vary significantly depending on the application, for example if you are classifying sleep stages then a window size of 1 sec to 10 sec is fine, as you don't expect someone to go from fully awake to sleepy in less than 1 sec. Similarly for drowsiness detection (transition from awake to drowsy is gradual), an EEG window size of a few seconds has also been utilized. This is an application driven parameter.

I would suggest you start with simple time-domain features first, especially something like "Hjorth parameters". Hjorth parameters extract some frequency-domain characteristics directly from the time-domain without having to use any frequency analysis tools like FFT or wavelet etc.. I have used Hjorth on both EMG and EEG and got excellent results in many cases.

Hi Daniel,

Kindly see the below papers,

Article A review of feature extraction and performance evaluation in...

https://www.sciencedirect.com/science/article/abs/pii/S0208521620300553

Regards,

Ali

Hi Daniel,

Yes, you can extract the same features from EMG/EEG signals.

On the other hand,

The first difference/divergence point is related to how you prepare your EMG/EEG signals for feature extraction. As you know, EEG is decomposed into several bands, these are: delta (30Hz) (each with a certain frequency range that varies between publications). You can extract any time-domain, frequency-domain, or time-frequency domains features either from the original EEG signal (with full spectrum, say 0--58Hz or whatever range), or from the individual spectral bands (after filtering to get that band of interest). In comparison to EMG, we usually extract the features from one band (10Hz to nearly 500Hz). Or, you can also partition the EMG spectrum into ranges and extract features from each range just like EEG (you will need to determine the EMG ranges for your task accordingly).

The other divergence parameters could be related to how you further clean the signals, with EEG you need to clean eye-blinks and/or jaw movements away and with EMG you need to clean ECG beats (depending on where you collected the signals from). There could be other factors to worry about (depending on your application).

Other divergence points between the two signals' feature extraction process could also be related to the windows size/other parameters. For example, with EMG we usually use 100 msec to 150 msec windows when working on prosthesis control and may be larger values for different applications (usually small values depending on the application). On the EEG side, this will vary significantly depending on the application, for example if you are classifying sleep stages then a window size of 1 sec to 10 sec is fine, as you don't expect someone to go from fully awake to sleepy in less than 1 sec. Similarly for drowsiness detection (transition from awake to drowsy is gradual), an EEG window size of a few seconds has also been utilized. This is an application driven parameter.

I would suggest you start with simple time-domain features first, especially something like "Hjorth parameters". Hjorth parameters extract some frequency-domain characteristics directly from the time-domain without having to use any frequency analysis tools like FFT or wavelet etc.. I have used Hjorth on both EMG and EEG and got excellent results in many cases.

Hi Daniel,

In line with Dr. Rami N. Khushaba’s comments, the same feature extraction methods can be applied to both EMG and EEG in order to project the signals to a lower dimensional space. In addition, I did investigate a range of time and frequency domain features for EEG signal characterization and found that a combination of four feature sets extracted from frequency domain (including: power spectrum ratio: PSR, median frequency MDF, mean peak frequency: mPKF, and the total power:TTP) yielded high and stable motor imagery movement intent decoding accuracies across subjects. A colleague also tested the same feature set for classification of phantom finger sensations evoked based on multi-channel EEG recordings, and the results were consistently stable and high. I have attached some of the articles published from the above studies.

If you think the features could be helpful in your study, you may reach out to me.

Good luck!