Preprocessing as a Pipeline

Level 2 turns the individual techniques of Level 1 into a pipeline that someone else can run. It runs from opening a file to shipping a per-subject quality-control report in eight lessons: reading every common format and writing the result out as EEG-BIDS; verifying a montage and flagging bad channels against objective criteria before interpolating them; choosing a reference with intent and predicting what it does to amplitudes and maps; setting filter cutoffs from the analysis goal, including the two-pass strategy that lets ICA see one version of the data and your analysis another; rejecting artifacts without letting the threshold become an analysis choice; correcting eye movements by regression and by ICA, and classifying components from converging evidence; knowing when artifact subspace reconstruction and its relatives are the right tool; and putting the whole thing in a configuration file with a seed, a log and a report.

Two threads run through the level. The first is rank: every interpolated channel, every average reference and every projector costs a degree of freedom, and the number has to be carried from the second lesson to the sixth or ICA will manufacture components out of nothing. The second is judgment made in advance: rejection criteria, component-removal policies and subject-exclusion rules are all defensible only if they were fixed before the effect was visible. The capstone asks for a script that takes raw BIDS to cleaned continuous data plus a QC report for a full cohort, unattended, with the per-condition rejection table in plain sight.