[Eeglablist] About band-pass filtering

[Marco Buiatti]

Dear Arno,

Arnaud Delorme wrote:

> Dear Marco,
>
> sorry for the late feedback.


in the middle of August my expectation was far worse than 2 days...

>
>> I have a question about filtering. I would like to band-pass filter 
>> some EEG data consisting of epochs of about 1.5 sec, sampled at 250 
>> Hz, between 0.5 Hz and 20 Hz. Since I have around 370 data points per 
>> epoch, I'm forced to use a 120-points order high-pass filter. As 
>> recomended many times in this mailing list, I first high-pass filter, 
>> then low-pass filter. But when I look at ERPs, I still see slow 
>> oscillations which seem suspicious to me.
>
>
>
> It seems hard to remove oscillations at 0.5 Hz (2-second period) in 
> 1.5-second epochs. You should apply the filter before epoching the 
> data. Another reason for filtering the continuous data is that filters 
> may introduce artifacts at the beginning and at the end of each data 
> epoch.
>
>> The problem gets worse if I filter the concatenated data before 
>> epoching. 
>
>
>
> Do you mean continuous data (or did you artificially concatenated your 
> data epochs)? 


I artificially concatenated my data epochs (because I do not have the 
continuous data at the moment). Filtering concatenated data epochs gives 
big problems at the boundaries. I agree that the right thing to do is to 
filter continuous data and then epoching.

>
>
>> Could this be due to the shortness of the highpass filter? I tried 
>> also with 1 Hz and 2 Hz, but the result is similar. Or anyone sees 
>> another explanation?
>
>
>
> You can visualize filtered data using the spectopo() function (menu 
> Plot > Channel spectra and maps). If you low pass data below 20Hz you 
> should see ripples above 20 Hz with about 20-40dB decrease in signal 
> amplitude. The same is also true for highpass. This way, you can 
> actually visualize which portion of the data was filtered out.


I followed your sensible advice, and I found out that:
a) low-pass filtering at 20 Hz works well, giving a homogeneous decrease 
of about 40 dB in about 6 Hz interval around 20 Hz;
b) high-pass filtering at 0.5 Hz has a very small effect of decreasing 
the power up to about 2 Hz. As you said, it is hard to remove 2-sec 
oscillations in 1.5 sec segments!

Thank you again,

Marco

-- 
Marco Buiatti

Neurophysique et Physiologie du Systeme Moteur - CNRS UMR 8119
UFR biomedicale Les Saints-Peres
45 rue des Saints-Peres 75270 Paris cedex 06
Tel: +33(0)142862146
Fax: +33(0)149279062
http://www.neurophys.biomedicale.univ-paris5.fr/~buiatti/