In 19 awake SEEG patients, filter choice caused most of the AR phase forecast error gap

Earlier results for autoregressive (AR) forecasting of SEEG phase reported errors of 86.9 deg and 60–63 deg on the same data and scoring. A 2^4 factorial ablation tested filter, window, AR order, and sampling rate on 95 reliable-peak pairs from all 19 subjects in OpenNeuro ds005624, using a 100-ms forecast horizon and an offline phase reference. The ablation reproduced the contrast: 85.0 deg for the causal FIR recipe and 58.4 deg for the zero-phase Butterworth recipe. The filter choice had the largest main effect (+16.1 deg; all 19 subjects; Holm p = 1.5e-5). Replacing the Butterworth with the FIR increased error by 16.6 deg from that recipe; band-width-only swaps changed error by less than 1 deg. All eight FIR cells had errors of 77.9–86.1 deg, near the causal Hann baseline of 81.1 deg. The native sampling rate, longer window, and AR order of 20 each reduced error by 2–5 deg, mainly with the Butterworth. The study used five reliable pairs per subject (95 of 958), 60 s per subject, and an offline Gaussian-band-pass reference that is not the true phase. It tested offline simulation only, not closed-loop stimulation.

SEEG · phase forecasting · autoregressive models · signal filtering · factorial ablation · OpenNeuro ds005624 · intracranial EEG