In 8 SEEG subjects, anti-aliased decimation to 200 Hz kept 999 Hz test AUC; 100 Hz and unfiltered sample dropping reduced it

We tested the lowest sampling rate that keeps a movement read-out in stereoelectroencephalography (SEEG). We used OpenNeuro ds004473 (8 subjects, cued key press, 102–107 trials per subject, 100–112 bipolar pairs). We read EDF records near each trial with HTTP byte ranges and reduced 999 Hz data by factors of 2 to 10 (500 to 100 Hz) with two front ends: an anti-alias finite impulse response (FIR) filter followed by decimation (F), and sample dropping without a filter (U). Features were log power in 13–30, 30–70, and 70–150 Hz, clipped at 0.9 times the Nyquist frequency. A shrinkage linear discriminant analysis (LDA) was trained on the first 60% of trials and tested on the last 40% (move: 0–1 s after Go Cue; rest: 1 s before Fixation). Native mean test AUC was 0.830. With F, AUC changes were +0.002 to −0.003 at 500, 333, 250, and 200 Hz (Wilcoxon p ≥ 0.25). Below 200 Hz, the 70–150 Hz band was lost and AUC decreased: −0.038 at 167 Hz, −0.047 at 125 Hz, and −0.060 at 100 Hz (7 of 8 subjects lower at 100 Hz, p = 0.039). With U, AUC was lower than native at all rates (−0.018 to −0.075; p = 0.016 at 200, 167, and 125 Hz; p = 0.0078 at 100 Hz) and lower than F at 250 Hz (p = 0.016) and 200 Hz (p = 0.047). Aliased high-gamma power did not replace the true band; it added noise. For this task, an offline 200 Hz read-out with an anti-alias filter used one fifth of the native data rate with no measured loss. Limits include one offline task, 8 subjects, no multiple-comparison correction, and decimation of data recorded at 999 Hz. The package includes code, per-subject AUCs, group tables, a figure, input hashes, and reproduction steps.

stereoelectroencephalography · SEEG · sampling rate · anti-alias filtering · decimation · movement decoding · band power · shrinkage LDA · AUC · OpenNeuro