In 12 SEEG patients, AR(20) had larger slow-oscillation phase errors than calibrated Hilbert hold at 100 ms

We tested whether autoregressive (AR) forward prediction reduces slow-oscillation (SO, 0.5–2 Hz) phase error at a 100 ms stimulation latency in mesial-temporal SEEG. The pre-specified primary set was six patients (sub-01–06); a replication set contained six new patients (sub-07–12). Each estimate used a 4 s causal window. We compared Yule–Walker AR(20), fit to the band-passed window after removing its last 250 ms, with a windowed Hilbert phase hold. Calibration used a constant per-channel circular-mean bias fit on the first half of decisions and scored on the second half. The primary hypothesis was not supported: AR+cal had larger error than E1-hold+cal in all six primary patients (57.6 vs 49.1 degrees; the pre-specified one-sided test for AR being better gave p = 1.0; the opposite-direction exact sign-flip test gave p = 0.016 and was post hoc). In the replication set, AR+cal was also worse in all six patients (62.2 vs 54.3 degrees; p = 0.016 for AR worse). At 100 ms, hit rates within 45 degrees were 0.490 vs 0.580 in the primary set and 0.454 vs 0.535 in replication. At 0 ms the calibrated methods were almost equal. From 50 to 250 ms, AR+cal was worse in all 12 patients at each latency. Exploratory AR orders 8 and 40 did not change the result. The E1-hold and E1-hold+cal results in the primary set (52.10 and 49.09 degrees) reproduced the prior internal result (52.1 and 49.1 degrees). The analysis was offline, used short 67–125 s segments, and defined truth with a non-causal Hilbert phase of the same band. Latency was a pure delay; real amplifier effects and stimulation artifacts were not simulated. Sleep stage was not scored, and only one AR design was tested. Six patients per set give a minimum exact p of 0.016; opposite-direction tests were post hoc.

slow oscillation · SEEG · phase estimation · phase prediction · autoregressive model · Hilbert transform · closed-loop stimulation · sleep · signal processing