I have a data set of EEG data acquired during the performance of multi-joint self-paced elbow extension movements, and I would like to know whether it is possible to quantify the amount of noisy information during the onset and further movement performance in channels C3, C4 and Cz, possibly produced by head or shoulder movement. I did not measure the head location or the EMG of the neck muscles during recording, so I only have the EEG data to do that.
I wonder if there is any measure, maybe based on power-spectrum or entropy to do that?
Hello, I rather work on multilead ECG, but the first idea which comes up to my mind is a blind source separation technique, e.g. PCA or ICA, so you can decompose your multivariate EEG signal in a set of components, and the last ones are generally represented by noise. The number of components can be determined by suitable criteria depending on the method. Have you tried anything similar?
Maybe you can try using passive movements, anyway EMG noise has much higher frequency than EEG...
Thanks to all anserws. I did think in these methods of blind source separation techniques. However, I was thinking in some method more time dependent, i.e., a time series of measure of contamination of the signal, that would allow to detect changes within each trial that could be associated with the onset of the movement.
Dear Julia,
Thank you for your question. I find it surprising that you would have (EMG) artifact in CZ/C3/C4. Essentially there is no muscle there. Do you mean electrode movement? I would be interested in seeing an example of a trace to see what artifact you are specifically referring to. Are you using a specific reference electrode that may be contaminated?
Very best regards
Eleftherios Papathanasiou,
Clinical Neurophysiologist,
Nicosia, Cyprus
I did not work on EEG signals. But I guess adaptive filters may help you for dealing with noise information.
To do a raw quantification of the amount of noise in your multivariate data, you can do a Vector Autorregressive model of your data and them do a state-space model of your system, via Kalman Filtering/Maximum Likelyhood and, then, estimate the variance-covariance matrix of your system.
The big problem of this approach is that you have to assume that your data is linear, stationary and Gaussian.
Rope that helps.
Regards.
But to do a first approximation, you can use this approach.
Regards.
Hi,
There is a good thesis allow you how to quantify SNR of EEG signals.
good luck. Reference of the Thesis provided below:
Naeini RY. Automatic Detection of High Frequency Oscillations of Neural Signals in Epileptic Patients. Concordia University, Montreal, Quebec, Canada, 2012 (Doctoral thesis).
We use trhe SNR (BrainVision software) to discriminate event-related power from the background activity at each sensor. May be you can use a similar aproach to isolate between movement-related power. I enclose our paper describing the method
As Francesco Di Russo pointed out already, EMG has a much broader spectrum than EEG. Hence, looking at the amount of energy in a band between, eg., 60 and 120 Hz may give an indication of EMG contamination.
If, however, the contamination is due to electrode movement (as suggested by Eleftherios Papathanasiou), it will show up as additive signal components at low frequencies (likely within the EEG band of interest), with one frequency component the pacing rhythm of your experiment. The amplitude of this type of contamination will depend on the electrode contact quality (impedance).
Thank's to all responses! I will try to work on it.
Happy new year to all
!
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing