12.10: Demonstration- Decoding in the Frequency-Domain

Although eStudio focuses mainly on time-domain ERPs rather than frequency-domain analyses, we’ve included a simple feature in our decoding tools that allows you to apply decoding to frequency-domain data. For example, the scalp distribution of alpha-band EEG oscillations varies according to where the participant is attending (Rihs et al., 2007), and the algorithms implemented in eStudio make it possible to decode the direction of attention from these oscillations (Bae & Luck, 2018).

To see how this is done, go to the EEG tab and open the Extract Bin-Epoched Single Trials (BEST) panel. If you don’t have any EEGsets loaded, everything in the panel will be grayed out, but that’s okay for now. If you look near the bottom of the panel, you will see a checkbox labeled Transform into phase-independent power along with two text boxes. When you are creating a BESTset from an EEGset, you can check this box and enter the desired frequency range in the two text boxes. To extract the alpha band, for example, you could specify 8 to 12 Hz as the frequency band. 

When you do this, each EEG epoch is passed through a bandpass filter with the specified lower and upper bounds, and then the filtered data are passed through the Hilbert transform before being saved in the BESTset. The filtering step extracts the desired frequency band, and the Hilbert transform converts the data from a voltage value at each time point to a phase-independent power value at each time point. Consequently, the single-trial epochs in the BESTset will represent the power in the desired frequency band at each time point. You can then just apply the regular decoding process to the resulting BESTset to decode the classes on the basis of the phase-independent power information. In other words, you just need to check the box labeled Transform into phase-independent power and specify the frequency range to switch from decoding voltage values to decoding in the frequency domain. It couldn't be much simpler! 

As a demonstration, I took the data from the EEG CORE N170 experiment and created new BESTsets with this option set to extract the phase-independent power from 8 to 12 Hz (you can find these BESTsets in the N170_BEST_Alpha folder). I then performed the decoding of faces versus cars (you can find the results in the N170_MVPC_Alpha folder). I’m not going to have you go through all the steps yourself (although you can if you want).

I did encounter one small problem the first time I tried to perform this analysis. As discussed in Chapter 7 of Luck (in preparation), you lose a lot of temporal information when you apply a narrow bandpass filter like the 8–12 Hz filter I used to extract alpha, and you need a relatively long epoch length to apply the filter. The epoch of -200 to 800 ms used in the ERP CORE N170 experiment was not long enough for the filter to work. I therefore re-epoched the data from -2000 to +2000 Hz prior to creating the BESTsets (you can find the re-epoched data in the N170_Data_Alpha folder). Also, when I performed the decoding using the Multivariate Pattern Classification panel, I clicked the Options button and set the time range for decoding to be -200 800 (to match the decoding time window in the previous exercise).

After I ran the decoding for each of the 9 participants, I made a grand average, which you can see in Screenshot 12.15. To tell the truth, I wasn’t quite sure what to expect. However, alpha is suppressed when you focus on an interesting visual object (as first demonstrated by my intellectual great great grandfather, Lord Adrian, in a paper by Adrian & Matthews, 1934). Faces tend to automatically capture attention, which presumably impacts alpha, so I thought the alpha signal might differentiate between faces and cars. As you can see in the grand average, the alpha signal could in fact be used to decode whether the stimulus was a face or a car, with a peak around 150 ms.

Note that the decoding accuracy first starts rising above chance at approximately 0 ms. This is a result of the temporal smearing that occurs when you isolate a specific frequency band, and it does not mean that the brain had information about whether the stimulus was a face or a car at that time point.

---