Epoch and trial rejection

The simplest artifact strategy: identify segments (epochs or trials) too contaminated to trust and exclude them from analysis. Criteria range from a fixed peak-to-peak or amplitude threshold to more robust, data-driven schemes (for example Autoreject, which learns per-channel rejection thresholds by cross-validation and decides whether to reject or interpolate each channel-epoch). Rejection makes no modelling assumptions and cannot distort the surviving data the way correction can, which is why ERP conventions still lean on it.

The costs are lost data and bias. Discarding trials shrinks the sample and lengthens sessions — costly for patients and for BCI calibration — and if artifacts correlate with the condition (blinks after targets, movement in one class), rejection biases the comparison. Rejection and correction are therefore complementary: correct systematic, time-locked artifacts such as blinks and reject the rare, unmodellable ones such as a large movement burst.