It is generally a bad idea to do channel selection individually for each subject because it can lead to type 1 errors (false positives) if you have a large number of electrodes. Also, it will make the interpretation of results more difficult if the selection process results in electrodes that are located distant to one another, since it implies that different brain processes are being used for the BCI/analysis, even if this is not the case.

Due to the poor spatial acuity of EEG, most nearby channels will display highly correlated activity. This means that it often a good idea to average together contiguous electrodes (ROI approach), which can control for the differences in head shape and cap placement across participants.

It also follows that if a channel is showing an effect, but its neighbor is not even close to showing the effect, then the second channel is likely contaminated with noise. It is therefore important to remove/interpolate bad channels before channel selection for main analysis. Of course, it may be that channels are on either side of a strong positive/negative current boundary, but this can be addressed by the ROI approach and by looking at the pattern of activity over the whole scalp. The size/location of the ROI can be chosen after looking at the data from all subjects. There is much less of an issue of bias by taking this approach, although it is arguably better to use a location based on previous studies.

I am assuming that you have a relatively high number of electrodes (e.g. > 32). If the number of electrodes is low, you may need to discard the participant on the grounds of poor signal quality.

Hi, Mohammad.

I see no problem with selecting individual features. In fact, given the large inter-subject variability, I would think this is a very reasonable approach. Recently, Seghier and Price have discussed the role of variability for neuroimaging applications (although not specifically for BCI tasks); it might interest you: doi 10.1016/j.tics.2018.03.003.

Moreover, individualizing subject's features for BCI applications and motor imagery tasks has, indeed, been done (e.g., doi: 10.1155/2019/3817124; 10.3390/s20061620; 10.1109/SMC.2017.8122964). Also, personally, in our research group, we have conducted several studies focused on motor imagery training (aiming BCI training), and results overall support individualized protocols.

Of course, it is also always helpful to consider the neurophysiological meaning of possible features, to analyze whether they do make sense for your application.

Regards,

Carlos.

You can use data-driven methods, such as CSP or similar techniques that practically optimize spatial filters for you. You can check out comparison I didArticle Effect of power feature covariance shift on BCI spatial-filt...