[Eeglablist] continuous data - why do ICA on continuous data

Arnaud Delorme arno at salk.edu
Thu Apr 7 11:03:06 PDT 2005

>Do you believe the reason 'better' results are obtained by doing ICA on
>continuous data rather than by doing ICA on concatenated epochs (that were
>previously segmented) is because of the discontinuities in the derivative of
>the signal at the concatenation boundries?
No this is unrelated to the discontinuities in the signal since time 
points (from all data epochs) are shuffled at each iteration in the ICA 
algorithms. From my perspective this is due to baseline removal: for 
instance if you have an eye movement in the baseline, you will subtract 
a different baseline from each data channel (since the amplitude ot the 
eye movement will be different in each channel). This introduces 
non-linearity in the data for this eye movement and ICA might have a 
hard time dealing with it (possibly dedicating a component to account 
for this specific eye movement). The same remark might be true for other 
types of artifacts that affect the baseline average potential. This is 
just a guess though. Also, the difference between ICA applied to 
continuous data and ICA applied to concatenated data epochs (when the 
baseline is 'clean') does not seem so dramatic and we are still using 
the later one in many occasions (for instance for consistency when we 
started this way for other subjects).


Note: do not use the SOBI ICA algorithm on concatenated data epochs. 
Since SOBI (unlike other ICA algorithms) uses time correlation, it will 
be affected by discontinuities in the data.

More information about the eeglablist mailing list