post SSP-SIR ICA artifact

[GiacomoBert]

In my pipeline I first suppress the TMS-evoked macular artifact with SSP-SIR and then refine the artifact rejection with fastICA.
I use SSP-SIR like this

EEG = pop_tesa_sspsir(EEG, 'leadfieldIn', [],...
            'leadfieldChansFile', [],...
            'artScale', 'automatic',...
            'timeRange', [])

when running fastICA with

EEG = pop_tesa_fastica( EEG, 'g','gauss','stabilization','on' )

and visualizing the components with

EEG = pop_tesa_compselect( EEG, ...
            'compCheck','on', ...
            'remove','on', ...
            'saveWeights','on', ...
            'figSize','large', ...
            'plotTimeX',[-50 100], ...
            'plotFreqX',[1 200], ...
            'freqScale','log', ...
            'tmsMuscle','on', ...
            'tmsMuscleThresh',8, ...
            'tmsMuscleWin',[8 50], ...
            'tmsMuscleFeedback','off', ...
            'blink','off', ...
            'blinkThresh',2.5, ...
            'blinkElecs',{'Fp1','Fp2'}, ...
            'blinkFeedback','off', ...
            'move','off', ...
            'moveThresh',2, ...
            'moveElecs',{'F7','F8'}, ...
            'moveFeedback','off', ...
            'muscle','off', ...
            'muscleThresh',-0.31, ...
            'muscleFreqIn',[7 70], ...
            'muscleFreqEx',[58 62], ...
            'muscleFeedback','off', ...
            'elecNoise','off', ...
            'elecNoiseThresh',4 ...
            ,'elecNoiseFeedback','off' );

I get this weird artifact component (I am uploading some examples). It seems to have always the same latency ~ -15 to 35 ms .

Is anyone else seeing this with their data?

[GiacomoBert]

Update: I think I found the source of this.

When using artScale = 'automatic', SSP-SIR estimates a time-varying weighting kernel based on the RMS of the >100 Hz filtered signal in a 50 ms sliding window. Because the window is 50 ms wide, samples at approximately -25 ms already "see" the TMS pulse energy entering the window. This means the SSP projection is partially applied to the last ~25 ms of the baseline, attenuating the signal there.

The result is a "pinch" in the pre-stimulus baseline from roughly -25 to 0 ms — the signal amplitude is reduced relative to the earlier baseline. When baseline correction is later applied over a window that includes this attenuated segment, the biased mean causes the pinched region to shift in the opposite direction, creating a visible pre-stimulus deflection.

I believe this is what ICA picks up as a component at ~-15 to 35 ms i.e. the transition zone where the sliding kernel ramps from near-zero to full weight, centered around the pulse. The latency is consistent across subjects because it's determined by the fixed 50 ms kernel width

For now I jsut switched from 'artScale', 'automatic' to 'artScale', 'manual' with an explicit timeRange that starts after the pulse (e.g. [5 50] or similar, depending on your interpolation window). this way the projection is controlled by a sigmoid weighting function that transitions sharply at the specified boundaries, so the baseline stays untouched. If you cut and interpolate the pulse artifact from e.g. -2 to 9 ms, a timeRange of [3.5 50] keeps the sigmoid ramp almost entirely within the interpolated segment.

Hope this helps!