which COEF values to keep in PRINCOMP for EEG feature dimension reduction ?

My EEG data contains 118 channels. Suppose for each trial I extract 4 features. Since all 118 channels have data for that trial, I have to extract 4 features from all those 118 channels.

So my feature matrix for each trial becomes 118*4 (channels*feature).

X is of size 118*4

[coef score latent]=princomp(X) %

Should I use princomp(X) or princomp(X')? If I use princomp(X), COEF is 4*4.

If LATENT shows that first 2 values have the most variance, should I use COEF(1:2,:) or COEF(:,1:2) as features? I mean first 2 rows or columns should I consider?

Will I use these 2*4 values as features instead of whole 118*4?

Please help