Causal AR prediction reduced 100-ms SEEG phase error, but median error remained 63° across 958 reliable pairs

We compared causal autoregressive (AR) forward prediction and linear phase extrapolation from a windowed filter (WIN-LIN) with one-sided Hann finite impulse response (FIR) estimators for phase prediction in awake stereo-electroencephalography. The evaluation used 958 reliable-peak bipolar pairs from 19 subjects in OpenNeuro ds005624 v1.0.0, using the same 60-s segments, peak frequencies, offline reference phase, amplitude gate, calibration/test split, and bias correction as the earlier FIR analysis. At a 100-ms horizon, pooled median absolute errors were 78.1° for FIR 4.5 cycles, 68.7° for FIR 3 cycles, 62.6° for AR(20), and 56.1° for WIN-LIN. AR(20) reduced subject-median error relative to FIR 4.5 cycles by 12.7°, with lower error in all 19 subjects (exact two-sided sign-flip p = 3.8×10⁻⁶); only 15.7% of AR pairs were below 45°. At a zero-ms horizon, AR and WIN-LIN errors were 28.6° and 31.5°, respectively, and 97.1% and 96.6% of pairs were below 45°. Error varied by frequency band. A synthetic 8-Hz phase-drift check gave AR errors of 4° without added noise and 9° at SNR 1. The findings show improved 100-ms prediction relative to the longer FIR baseline, but prediction remained poor for many pairs. The offline reference uses the same Gaussian filter shape as the window methods and may favor them; data comprise one 60-s segment per subject, no stimulation was tested, and bias was calibrated.

stereo-electroencephalography · phase prediction · autoregressive prediction · causal filtering · signal processing · neural oscillations