Automated rejection and repair of bad trials in MEG/EEG

Abstract

We present an automated solution for detecting bad trials in magneto-/electroencephalography (M/EEG). Bad trials are commonly identified using peak-to-peak rejection thresholds that are set manually. This work proposes a solution to determine them automatically using cross-validation. We show that automatically selected rejection thresholds perform at par with manual thresholds, which can save hours of visual data inspection. We then use this automated approach to learn a sensor-specific rejection threshold. Finally, we use this approach to remove trials with finer precision and/or partially repair them using interpolation.We illustrate the performance on three public datasets. The method clearly performs better than a competitive benchmark on a 19-subject Faces dataset.