Anti-aliased decimation to 250 Hz preserved key-press decoding in eight sEEG subjects

We tested how lower sampling rates affect offline decoding of a cued key press versus pre-fixation rest in stereo-EEG (sEEG) recordings from eight subjects in OpenNeuro ds004473. Each subject contributed 160 trials. We extracted log band-power features from adjacent bipolar contacts and evaluated a Ledoit-Wolf shrinkage linear discriminant using a time-ordered 60/40 train/test split. We simulated decimation from approximately 1000 Hz, with and without a 201-tap anti-alias filter. In the pre-registered one-sided sign-flip test with Holm correction over three tests, 250 Hz with filtering was non-inferior to native rate at a 0.02 AUC margin: mean paired AUC change was -0.0003, all eight subjects were above the margin, and adjusted p = 0.012. At 200 Hz, non-inferiority held at p = 0.023 unadjusted; it failed at 125 and 100 Hz. The mean AUC change at 100 Hz with filtering was -0.009 and was not significant (Holm p = 0.68). Unfiltered sample dropping performed worse than filtering at every rate, with mean change -0.050 at 100 Hz (Holm p = 0.93 for the pre-registered comparison). The data were fetched from an unversioned OpenNeuro S3 mirror; although the preregistration names ds004473 v1.0.2, the exact source release cannot be confirmed. This is a simulation using clinical-amplifier recordings, one task, one train/test split, and band-power features up to 150 Hz; it does not establish performance for a low-rate implant or features such as ripples or spikes.

stereo-EEG · sampling rate · anti-alias filtering · decimation · key-press decoding · AUC · band power · OpenNeuro ds004473