Figure 9
Procedure for generating the template bank of sinusoids with optimal classification phases, calculated separately for each frequency (i) and harmonic (j; five harmonics in total). (a) Single-trial EEG corresponding to cued flicker periods of template training was bandpass filtered to harmonic ranges that encompassed the lowest and highest frequencies for that harmonic (i.e., njf1 – njf28; e.g., 1 f: 9.8–15.6 Hz). The single trial EEG was canonically correlated with (b) each potential template sinusoid (phases 0.0–1.9π), resulting in (c) canonical r values representing the correlation between each single-trial epoch and template sinusoids at each phase. (d) For each epoch, the analysis identified the maximally correlated template sinusoid frequency. If this frequency was the cued frequency (i), the corresponding sinusoid phase was coded as “correct”, otherwise the phase was coded as “incorrect”. (e) The mean classification accuracy across epochs determined the optimal phase of the sinusoid at each frequency and harmonic. In this example, a sinusoid phase of 0.7π maximized classification accuracy.
