[FieldTrip] Decreased baseline level after using ICA in ERP data

Matt Craddock matt.craddock at uni-leipzig.de
Tue Jun 25 17:17:43 CEST 2013

Dear Beom Jun,

Regarding detrending - ICA works better with relatively stationary data, 
which is why high-pass filtering - as Diego mentioned - is often 
performed. Both detrending and high-pass filtering remove/attenuate slow 
fluctuations in the signal, so I'd suggest using one or the other 
procedure before running ICA if you think such low frequency activity is 
affecting your decompositions.


On 25/06/2013 13:04, Beom Jun Min wrote:
> Dear Diego,
> Thank you for your kind answer.
> The importance of 'quality' you mentioned and the references that you
> attached could help me to understand the ICA algorithm further.
> And I have an additional question about preprocessing before ICA. Is
> detrending needed before the decomposition if there is a linear trend in
> the segmented data?
> Because I noticed one component showing linearly and consistently
> decreased (or increased) activity during one segment in some trials
> after ICA, I wondered why that happened.
> Apart from that, I found the possible cause of the past problem. It
> looks like the ft_rejectcomponent might remove the 'demean' effect.
> After I used the function without any component removing, (cfg.component
> = [];) the baseline level decreased again but the shape of the ERP does
> not change.
> However, I have not found the way to correct this decreased baseline
> yet. The ft_preprocessing with demean pre-stimulus does not work.
> Thanks.
> BJ

Dr. Matt Craddock

Post-doctoral researcher,
Institute of Psychology,
University of Leipzig,
Neumarkt 9-19,
04109 Leipzig, Germany
Phone: +49 341 973 95 44

More information about the fieldtrip mailing list