IRASA- periodic power values

[apoorva6262]

Hi @raphaelvallat ,

I used the following function to determine the periodic power from the signal : freqs, psd_aperiodic, psd_osc, fit_params = yasa.irasa(sig, Fs, band=[3,60],kwargs_welch={'average': 'median'})

This gave me the output for average periodic power (np.nanmean(psd_osc)) as 4.913666773031852e-14.

I tried neurodsp package (https://neurodsp-tools.github.io/neurodsp/auto_tutorials/aperiodic/plot_IRASA.html?highlight=irasa) with same settings and I got the average periodic power to be 1.3315638777465248e-13. Wondering what could cause the difference? I used the same welch and median method to determine the psd.

Additionally how do you determine the periodic power for a subset of band frequencies like alpha from the initial fit ? Right now I have set the band parameter to 3-60hz, but how do I extract alpha frequencies from the fit itself ? I did not want to set the initial band paramaters to 8,12hz as that is a small range to find the aperiodic fit. I tried indexing but that gave me some shape issue.

Thanks, Apoorva.

[raphaelvallat]

Hi @apoorva6262,

Regarding your first question, these two values are actually both zero:

np.isclose(4.913666773031852e-14, 1.3315638777465248e-13)  # True
np.isclose(4.913666773031852e-14, 0)  # Also true

For your second question, I would recommend using the yasa.bandpower_from_psd function (https://raphaelvallat.com/yasa/build/html/generated/yasa.bandpower_from_psd.html):

yasa.bandpower_from_psd(psd_osc, freqs, band=[(8, 12, "Alpha")])

Hope this helps, Raphael

[apoorva6262]

@raphaelvallat , For the second question, I replaced psd for 'psd_osc' in 'yasa.bandpower_from_psd' as I am interested in periodic peaks only :

freqs, psd_aperiodic, psd_osc, fit_params = yasa.irasa(sig, Fs, band=[3,60],kwargs_welch={'average': 'median'})
yasa.bandpower_from_psd(psd_osc, freqs, ch_names=['F7'], bands=[ 4, 8, 'Theta', 8, 12, 'Alpha',  13, 30, 
                                                         'Beta'], relative=True) 

The shape of psd_osc is (1,229) , freqs is (229), both are numpy arrays. But I get this error: TypeError: 'int' object is not subscriptable

[raphaelvallat]

Please try the following:

yasa.bandpower_from_psd(psd_osc, freqs, ch_names=['F7'], bands=[(4, 8, 'Theta'), (8, 12, 'Alpha'), (13, 30,  'Beta')], relative=True) 

[apoorva6262]

Perfect thanks! One last question, a lot of the PSD values are negative and I did see your post on how to handle it. But I was wondering if you had come up with any other solution besides the ones you have mentioned here https://github.com/raphaelvallat/yasa/issues/29 ?

[raphaelvallat]

Unfortunately I do not have new solutions to this problem :/ Have you tried using FOOOF? It is also a great alternative to IRASA.

[apoorva6262]

No worries, I did try FOOOF and that is great. I am actually comparing IRASA to FOOOF

[apoorva6262]

@raphaelvallat , Another question, I am running this IRASA function to determine periodic power. For some files, I see that the fit(r2 value) is poor. Do you have any suggestions as to how to exclude poor fits or any outlier methods you employ for further statistical analysis?

[raphaelvallat]

I have not used IRASA in a long time for my own analyses so it's hard to say. I would probably not exclude a night just based on a poor R^2 value. But I do recommend plotting the full, aperiodic, and periodic spectra to get a better idea of what may be happening.

[apoorva6262]

Thanks! With my dataset, I am determining power across all 32 eeg channels for three different groups. For consistency sake, I have made sure to use the same settings of IRASA for all the channels and for all subjects in different groups. But there are 1-2 subjects with poor fits due to noise in higher frequencies (r2=0.002) , would you recommend just reporting the noise? It wouldn't make sense to use different settings for each file/subject to get a better fit.

[raphaelvallat]

Can you try to filter out the high-freq noise for these 1-2 participants? If the power spectrum is really too noisy then I don't think it's worth calculating IRASA honestly, the results will likely be biased.

[apoorva6262]

Also , I have this issue of negative psd issue with vast majority of subjects, i have used the default resampling method.