Baseline normalization of time-frequency power

[@SUD236]

Hi Mike! It's guiding to see your lectures and implement them. I just want an expert comment on my pipeline. Your comment might mark some subtle or major conceptual errors. The pipeline is as follows: 1. Event Extraction (Single Trial) 2. Bandpass filtering 4.0-40.0 Hz 3. Re-referencing (Average) 4. Correlation and thresholding based bad channel detection and interpolation 5. ICA computation, Automatic Component Labeling, and Rejection of components with artifacts namely eye, muscle, heart, channel noise, and line noise. 6. Baseline normalization (Decibel Conversion) to counter the gradual change in impedance (EGI EEG System with saline solution), subject variation, 1/f phenomena (though it has been taken care of by filtering itself), approx normal distribution of normalized data. Please comment on this.

[@mikexcohen1]

A high-pass filter at 4 Hz seems a bit excessive, though it depends on what you're doing with the signal. I would also be reluctant to trust an automatic IC rejection for the features you list. I recommend following my playlists from the beginning; several of these points are discussed in earlier videos.

[@SUD236]

@@mikexcohen1 Will surely do in order. Thanks very much for your reply.

[@claudiakrogmeier557]

@@mikexcohen1 In my pipeline, I was planning to do a bandpass filter between 1-30Hz, or between 2-30 hz, but I have also seen .5Hz-40 or .1-30 Hz. How much do these different filters affect the data? I am only interested in the alpha range, (which I have seen as 9-11, 8-12, and also 8-13 Hz, and am not sure why they are different). I have not found justifications for specific filtering frequencies. I would appreciate your input, thank you so much!

[@SUD236]

Hi Mike! I am confused about the order in which I should put the baseline normalization in my pipeline if I am performing non-linear normalization (say, decibel normalization). Should I end up doing the baseline normalization at last (given the slow wave drifting is already taken care of)? Please add your valuable comment.

[@mikexcohen1]

Yes, baseline normalization would be the last step, after averaging the raw power over trials.

[@haneensuradi]

Hello Mike! Thanks for this very useful lecture. I went through the code you provided in your course on the baseline normalization. I found that after convolution, you average over trials and then you do baseline normalization on the averaged version of the data. Is it possible to do the baseline normalization on single trials without averaging? And if so, do we consider our baseline as the baseline of the average or the baseline of each trial separately? *I am talking here about condition specific baseline*

[@mikexcohen1]

I do not recommend single-trial baseline. It tends to be unstable and give unpredictable results.

[@claudiakrogmeier557]

Would temporal smearing result from an FFT on the data, or is temporal smearing specific to TF analysis (as opposed to simple power spectrum extraction) ? Thank you !

[@mikexcohen1]

Well, the FFT shows the spectrum across the entire window, meaning it's not possible to know *when* a particular narrowband even occurred, just from looking at the power spectrum. That's the primary motivation for a TF analysis. The smearing comes from having overlapping time windows, e.g., one time window from 0-500ms and the next time window from 50-550ms.

[@Academixx]

Hi Dr Mike, If i have a single task with 30 subjects involved (single trial), is it right to take the baseline averaged from all individual baseline?

[@mikexcohen1]

No, that's probably not a good idea. One of the motivations for using a baseline normalization is because of individual differences.

[@Academixx]

@@mikexcohen1 Tq very much sir.