Causal slow-oscillation triggers had 44° phase error and retained 13% of oracle HFA gain in an offline 70-subject SEEG study

We evaluated a causal slow-oscillation (SO, 0.5–2 Hz) phase estimator and simulated phase-triggered high-frequency activity (HFA, 70–150 Hz) in 70 subjects with NREM sleep SEEG (515 bipolar pairs), using 60 seconds per subject from three OpenNeuro datasets. The causal E1 estimator, applied to a trailing 4-second window, had a median absolute phase error of 44.0° (95% CI 42.7–44.9) and a median causal-minus-reference phase bias of +16.4° (positive in 68/70 subjects). In a calibration/test split within each 60-second segment, E1 triggers at each pair’s calibrated up-state phase produced a mean HFA gain of 0.026 z at zero delay. The planned test that this gain exceeded zero was not supported (Wilcoxon p = 0.10; sign-flip p = 0.10; 39/68 subjects positive). Oracle triggers produced a mean gain of 0.199 z, so E1 retained 13% of oracle gain (95% CI for the ratio −0.08 to 0.30). E1 triggers nevertheless exceeded anti-phase triggers (p = 1.5×10⁻⁴), and this difference persisted at 50 and 100 ms delays. Post hoc, the mean E1 trigger-locked HFA peak occurred about 110 ms before the trigger, consistent with triggers arriving late at the HFA level; this was not tested as a mechanism. The findings replicate the reported phase error in a larger independent dataset but show that phase error alone did not ensure up-state targeting. This was an offline simulation: the reference used the same band filter, the calibration and test halves came from the same segment, and there was no sleep staging, interictal-discharge removal, or stimulation.

slow oscillation · SEEG · causal phase estimation · high-frequency activity · closed-loop stimulation · offline simulation