Zero-filled gaps in OpenNeuro ds003374 EDF recordings bias high-gamma pre-onset baselines by a median 8%

We examined nine macro-contact EDF recordings from OpenNeuro ds003374 v1.1.1 (amygdala SEEG; 17 aversive or neutral film blocks of 24 s). Each 522-s, 2000-Hz recording contains 16 exactly zero-filled 5-s gaps in all 28 channels. Each gap begins at a block end and ends 2.000 s before the next block onset, so the files join 2-s pre-onset and 24-s block epochs with zero fill rather than continuous data. We compared fourth-order zero-phase Butterworth filtering of the full record (W) with filtering each non-zero segment alone (S). For 70–145 Hz high-gamma power, W raised power in the 2-s pre-onset window by a median 7.8% across channels (maximum 41%); most of this increase occurred in the first 50 ms after the gap. The resulting artificial onset change (0–2 s after onset versus 2–0 s before) had a median W-minus-S bias of −8.2% across subjects (range −2.8% to −12.5%; all nine subjects negative). For 1–30 Hz signals, the pre-onset difference was below 2%, and the per-subject onset bias was within ±0.009 log10. Subjects with equal NIX file sizes (05/08 and 06/09) were different recordings: the maximum cross-subject channel |r| across 36 subject pairs was 0.038, and NIX SHA-256 values differed. The equal-sample fraction of 0.153 in every pair was caused by shared zero gaps. Processing epochs separately or masking gaps before filtering avoids the identified transient. This analysis covers macro EDF only, tests one filter type, does not determine the real inter-block intervals, and does not establish the pipelines used in earlier publications; no inferential test was performed because the effect is deterministic for a given pipeline.

intracranial EEG · amygdala · zero-filled gaps · filter transients · high-gamma power · pre-onset baseline · OpenNeuro ds003374 · signal processing