Neurophysiological screening of individual variability for robust decoding in c-VEP-based BCI

Code-modulated visual evoked-potential (c-VEP)-based reactive brain-computer interfaces (BCIs) deliver high information-transfer rates with minimal calibration, yet performance often collapses when models are transferred between users. We, therefore, pursue a two-fold aim: first, to pinpoint neurophysiological predictors that explain this inter-participant variability; second, to identify a decoding pipeline that sustains accuracy across users in a burst-c-VEP paradigm (brief, aperiodic flashes at 3 Hz). From 24 participants, we find that stronger inter-epoch correlation ( $R\approx 0.80$ ), larger peak-to-peak amplitude of the flash-VEP, larger $\alpha$ bandpower, larger $\theta$ bandpower, and lower $\delta$ bandpower are five neurophysiological predictors that correlate between high performers ( $>$ 90% accuracy) and low performers ( $<$ 70%), enabling a 22 $s$ "go/no-go" calibration. We then compare three preprocessing schemes (small, combined, participant-specific) paired with three decoders-a convolutional neural network, a Riemannian xDAWN-LDA baseline, and GREEN, a wavelet-based symmetric positive definite neural network. Subject-specific alignment plus GREEN achieves 93% trial-level accuracy in both intra- and cross-participant settings, eliminating the 15-20% transfer loss obtained with the other tested decoding models while keeping the total calibration under 1 min. In conclusion, rapid user screening with these neurophysiological predictors, followed by this lightweight, user-specific pipeline, yields burst-c-VEP control that is fast to deploy and robust across individuals.