CCA and TRCA Spatial Filters (SSVEP)

Steady-state visual evoked potential BCIs need a spatial filter that pools channels to detect which flicker frequency the user is attending. Canonical correlation analysis (CCA) does this without any training data: it finds the channel weights (and reference weights) that maximize the correlation between the multichannel EEG and a reference set of sine and cosine waves at each candidate stimulus frequency and its harmonics, and the frequency giving the largest canonical correlation is selected. It is robust and simple but ignores subject-specific response shape.

Task-related component analysis (TRCA) improves on this by learning, from a few calibration trials per target, the spatial filter that maximizes reproducibility of the response across repetitions of the same stimulus, then matching new data against per-target templates. TRCA and its ensemble variant are the basis of the highest-throughput SSVEP spellers, reaching well over 100 bits per minute in trained users, and illustrate the general shift from stimulus-locked reference models to data-driven, subject-specific templates.