Autoregressive prediction reduced 100-ms phase error in sampled awake SEEG pairs, but its advantage depended on implementation

Using 60-s segments from 19 awake SEEG subjects in OpenNeuro ds005624, this analysis compared autoregressive (AR) forward prediction with causal Hann FIR estimators for phase 100 ms ahead of an offline zero-phase reference. It scored a hash-selected sample of 228 reliable-peak and 152 other bipolar pairs on a grid every 50 samples; calibration and test each used 27 s. For reliable-peak pairs, pooled median error was 59.8° for AR, 73.1° for the 3-cycle Hann FIR, and 79.7° for the 4.5-cycle Hann FIR. AR had lower subject-median error than the 3-cycle estimator in 18/19 subjects (median paired difference −11.1°, exact one-sided sign-flip p = 1.1×10⁻⁵); 20.2% of AR pair errors were below 45°, compared with 5.3% for the 3-cycle FIR. For other pairs, pooled medians were 76.8° for AR and 81.4° for the 3-cycle FIR. Above 13 Hz all methods were near 80°. The result challenges a prior report of near-chance AR performance on overlapping data, but the implementations differed and the cause was not tested. In a synthetic narrow-band check, AR was slightly worse than both FIR estimators. Findings are limited by the sampled pairs, coarse evaluation grid, one 60-s segment per subject, an offline-filter reference rather than true phase, and the absence of hardware latency or real-time speed testing.

autoregressive prediction · phase prediction · stereo-EEG · SEEG · causal filtering · Hann FIR · OpenNeuro ds005624