[Eeglablist] Extracting ASR weights and applying them later

[Makoto Miyakoshi]

Dear Andraz,

> If I understand correctly, the cutoff of the high-pass filter does not
have much influence on the ASR results?

It does.
ASR is sensitive only to amplitude. And the lowest frequency of the EEG
data typically have the highest power (unless your data shows really
overwhelming alpha peak, which is not rare). Therefore, if the input data
to ASR is not well high-pass filtered, ASR will correct more data, and
clean_rawdata() will reject more windows in the end. This is why the (IIR)
high-pass filter is equipped in the very beginning of the clean_rawdata()
suite.

> In your paper in Epilepsy Research, you found that ICA or ASR+ICA reduces
phase-amplitude coupling (PAC). Do you suggest computing PAC on minimally
preprocessed data (i.e., only high-pass filter, cleanline, channel
interpolation)?

Any signal processing is to cut both signal and noise, hoping to cut more
noises than signals to improve SNR. So, as long as SNR improves, cutting
some amount of signals should be acceptable as necessary evil. If you can't
accept this principle, you don't want to use noise reduction processes. So
my answer is, try both and see which one is more satisfying to you. At
least you have choices.

> Are you aware of any other work that investigates the effect of cleaning
procedures on PAC?

I know there are other tribes that have different ways to 'clean' EEG data.
Some of them seem reasonable to me, some are not. I can't speak of the
methods I haven't used.
But, generally speaking--for example, someone told me that using
wavelet-based data cleaning is good for data cleaning. But, good, in terms
of what? For me, the more critical question is what determines the
rejection criterion. Selection of the linear space (spatial filter vs.
frequency filter) is less critical. When to declare something as a noise is
a domain problem specific to EEG research. In that sense, ICA+ICLabel is
one of the reasonable solutions available today that can be used in
automated preprocessing.

Makoto


On Wed, Aug 30, 2023 at 6:49 AM Andraž Matkovič <andraz.matkovic at gmail.com>
wrote:

> Thanks for the explanation. If I understand correctly, the cutoff of the
> high-pass filter does not have much influence on the ASR results?
>
> Another related question: In your paper in Epilepsy Research, you found
> that ICA or ASR+ICA reduces phase-amplitude coupling (PAC). Do you suggest
> computing PAC on minimally preprocessed data (i.e., only high-pass filter,
> cleanline, channel interpolation)? Are you aware of any other work that
> investigates the effect of cleaning procedures on PAC?
>
> Best,
> Andraž
>
> V V sre., 30. avg. 2023 ob 01:06 je oseba Makoto Miyakoshi via eeglablist <
> eeglablist at sccn.ucsd.edu> napisala:
>
>> Dear Andraz,
>>
>> The short answer is no.
>>
>> If any, if the input data's PSD is similar to the preemphasis filter i.e.
>> the 8th order Yule-Walker function that was heuristically determined by
>> Christian (the developer) as an inverse PSD of human EEG data (i.e. the
>> flattening filter, so to say), then ASR's artifact detection (according to
>> its own design) works most effectively. But this does not mean it helps
>> the
>> algorithm per se.
>>
>> > This is useful because the ICA solution can be improved with certain
>> preprocessing steps (high pass filtering at 1-2 Hz), but these
>> preprocessing steps don't necessarily have to be applied to the dataset to
>> which the ICA weights are applied.
>>
>> I have words of warning for you. See my reply to Jan. Basically, I do not
>> recommend it. Remember, ICA's sensitivity is not uniform across all the
>> frequencies, but its sensitivity is (practically) proportional to the
>> amplitude. See my preliminary results from here
>>
>> https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline#Why_does_IC_rejection_increase_gamma_power.2C_or_why_is_an_IC_not_broadband-independent.3F_.28For_160.2C000_page_views.2C_05.2F10.2F2021_added.2C_06.2F27.2F2022_updated.29
>>
>> Would you be surprised to hear that IC rejection generally INCREASES gamma
>> power? If you are surprised, read my Wiki article above. It is partly
>> related to what you are asking. Importantly, the lowest end of the power
>> spectrum usually has highest power, that's the point. If you are omitting
>> 50 Hz and above for running ICA, practically it'll do no harm.
>>
>> Makoto
>>
>> On Tue, Aug 29, 2023 at 1:02 PM Andraž Matkovič via eeglablist <
>> eeglablist at sccn.ucsd.edu> wrote:
>>
>> > Dear all,
>> > I have a question about ASR (Artifact Subspace Reconstruction). In ICA,
>> it
>> > is possible to extract ICA weights and apply them to another data set.
>> This
>> > is useful because the ICA solution can be improved with certain
>> > preprocessing steps (high pass filtering at 1-2 Hz), but these
>> > preprocessing steps don't necessarily have to be applied to the dataset
>> to
>> > which the ICA weights are applied. For example, I can run ICA with a 1
>> Hz
>> > high-pass filter, but apply ICA weights to data with a 0.1 Hz high-pass
>> > filter. I am wondering if this is possible with ASR.
>> >
>> > Best regards,
>> > Andraž Matkovič
>> > _______________________________________________
>> > Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> > To unsubscribe, send an empty email to
>> > eeglablist-unsubscribe at sccn.ucsd.edu
>> > For digest mode, send an email with the subject "set digest mime" to
>> > eeglablist-request at sccn.ucsd.edu
>> _______________________________________________
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> To unsubscribe, send an empty email to
>> eeglablist-unsubscribe at sccn.ucsd.edu
>> For digest mode, send an email with the subject "set digest mime" to
>> eeglablist-request at sccn.ucsd.edu
>
>