The multi-taper method

[@ananthakrishnank3208]

Explained in a lucid manner. Thanks a lot!

[@wavesc89]

Great video, as usual! Could you explain (or point me to where it's explained) why this MTM is recommended for non-time nor phase locked signals and not for the other cathegories? Thank you :)

[@ananthakrishnank3208]

I wish to implement ML on EEG, for the two cases - with and without Multi tapering. Previously I fed raw EEG to 1D CNN. But to see the efficiency of Multi tapering method, do I have to feed spectrogram (with and without MT) to CNN?

[@mikexcohen1]

Yes, you'd use a 2D CNN if you're inputting the spectrogram.

[@ananthakrishnank3208]

@@mikexcohen1 The problem is I do find a function that return all three values required for spectrogram plotting - times, frequencies, PSDs. However, I do not see any readymade function that returns all three with Multi tapering. The best I could find is "psd_array_multitaper()" from mne library. But it returns only two values - frequencies and PSDs. So I understand that we will have to segment the EEG signals (say, each segment is 300 ms) and apply Multi tapering to each. Now each segment corresponds to middle time value. I mean, a segment from 0 to 300 ms corresponds to t= 150 ms. And a segment from 300 to 600 ms corresponds to t= 450 ms and so on. This way we also got the time axis. But the problem that I am facing is that psd_array_multitaper returns a different set of frequencies for each segment, hence I could not easily concatenate them column wise and get the spectrogram. What is the way to get the spectrogram with MT applied?

[@ananthakrishnank3208]

@@mikexcohen1 Since I have a 14-channel device, is it normal to concatenate 14 spectrograms into one spectrogram, before feeding it to a 2D CNN?

[@sulemanrasheed1634]

@@ananthakrishnank3208 my suggestion would be not to average activity across all channels as some activity might be channel-specific that will be lost in averaging Rather you can concatenate spectrograms so each input to ML model will be of shape (time, freqs, chs) etc and later you can apply CNNs just like you apply them to RGB images with first layer kernel depth to be 14 instead of 3. Good luck

[@mingxingli1640]

Thank you so much!

[@zilingluo1651]

Thank you very much ！！This video helps a lot!

[@marchleslie3648]

Excellent!