50Hz artefact for time frequency analysis

Hi all,

Thanks for your reply,

Indeed the artefacts are coming from the cHPI (I did not say that it is not
picking up the electrical mainline)

But here is a snapshot (see files in the link below) of one participant/part
with the artefact

You have view of few electrodes and this is raw data

The green line correspond to the start of the cHPI

You could see that before some electrodes were bad and remain unchanged

However some EEG were good before and with cHPI they have picked up some
high frequency

Also In general you could see that all EEG picked up some level of artefacts

I gave you also a plot for this subject/part of frequency decomposition

The EEG have 50z, 150z & around 330Hz

See files in art2/art4

https://www.dropbox.com/sh/hm7pdbhoc98yxf2/AABlCIdmLs4sWHw3W0sHesYGa?dl=0

Elekta is aware about this problem as it was recurrent in our recent
recordings and they are investigating

For the analysis,

I started to use a notch filter and it seems that it is removing the
artefacts as I would like (filtering the 3 sensors together)

However by comparing the unfilter vs notch filter data

I noticed that the filter data have another low frequency component added
(around 10Z) , effect present in the Mag

See figure art5 in the same dropbox..

I am using MNE_python for this notch filter with the command

raw.notch_filter(np.arange(50, 251, 50), method='fft', n_jobs=4)

any tips for this low frequency band addition?

Thanks for your answers

Elisabeth

Hi Elisabeth,

As someone else pointed out, your 50 Hz is almost certainly from the mains power.
That is almost surely the source of the 150 Hz also.
Noise from this source have the properties that (1) the frequency is very stable and (2) the amplitude is almost always also very stable.
Because of these properties, this type of noise it typically best removed by measuring its amplitude and phase over a fairly long period, say a second or more, and then subtracting it off. This approach is typically very precise at remove the line noise while leaving any 50 and 150 Hz activity which is not constant in the signal.
One way it can be done is by (1) performing a Fourier transform, (2) attenuating the coefficients of the two frequencies, and (3) reconstructing the modified signal with an inverse Fourier transform.
This is likely a better approach than convolution with a notch filter since it explicitly uses the "constant amplitude" property of this noise.
If the tools are readily available in mne-python, it might be worthwhile to try before chasing the low frequency noise you mentioned.

One other thing to consider is whether it is actually necessary to remove the noise at all.
If the method you plan to use to explore the gamma band activity has high spectral resolution, the line noise may just show up as sharp peaks which you can ignore.
On the other hand, if your method has low spectral resolution, it definitely is worthwhile and likely it will make sense to use some high resolution spectral method to assess whether the spectrum is adequately smooth across the two line noise peaks.

Finally it might be worthwhile to check a few data samples from other runs to see if line noise is consistently showing up on those same channels. If so, it's likely that the noise pickup is due to loss of proper "balance" by the amplifiers or to degradation in the connections which run to those amplifiers.

Regards,

Don

I'd preserve the alpha peak response (~10Hz) providing it is present (and
to the same extent) in the non-cHPI part of your dataset. /V