I have written a procedure to pre-process combined TMS-EEG data, using independent component analysis to reduce the extent of the tms artefact in the eeg signals. I have been able to reduce the noise in the data, though I am convinced that this could be done more efficiently and more effectively. Is it possible to apply a method that can quantify which signals are most likely to be strong representations of the artefact. Although this would not act as a substitute for visual inspection and studying of the components it would be a useful starting point. Thanks.
Thanks for the suggestion Oliver. Sadly PCA wouldn't be very useful in this case. What I have done is to break down the data into it's independent components, and am now searching for a way of quantifying which are the strongest representation of the artefact. I've attempted a cross correlation between the experimental data and the sham trial data but that didn't prove too helpful. Stuck for ideas now.
try a perceptron neural network with a Features Selection. That should give you the most important components implied in your signal with a way to calculate the impact (or weight) of each of them. Check this paper out :
Article Features Selection and Architecture Optimization in Connecti...
do you want to use the ICA to remove the TMS artifact or to improve the signal after you remove the artifact with after methods?
I don't believe using ica really help in the first case. It will likely remove part of the magnetic artifact but you'll also remove most of the brain activity around the stimulus and this is quite evident if you look at the data from the electrodes surrounding the hot spot...and you'll still have artifact.
It is not possible to "remove" the artefact entirely, however, the methods I have been implementing are based on techniques currently in use at the Illmoniemi lab. I agree that you stand to lose some cortical activity if not careful, hence the search for a method of quantifying which components are the strongest representation of the artefact. I have had a look at the TMS-EEG paper that you have on your contributions, this seems quite similar to the approach in Taylor et al, 2010. However, by leaving in the components that add to the decay and muscle artefacts the data shows a distortion that adds an inherent difficulty to the interpretation of the signals. The need here is to reduce the effects of the artefact as a whole (so the 20-50ms that it occurs over), so reducing the slow wave components and minimising muscular activity - but not actually taking out the pulse. This should be possible by removing up to 10/15 components from a high density set up of 128 channels/components.
It depends on the system you use to acquire the eeg...with the system we use there is no long decay of the signal, so no more than 7 ms to trash.
- Badges
- Science topic