A calibrated constant offset matched frequency-advanced ecHT for 100 ms SO-phase prediction in held-out SEEG from 6 patients
We simulated a 100 ms latency between the phase decision and pulse for slow-oscillation (0.5-2 Hz) phase-locked stimulation. We used new, non-overlapping data from OpenNeuro ds003498 v1.0.1 (6 patients, 302 bipolar mesial-temporal SEEG pairs, interictal NREM sleep). We compared a windowed Hilbert estimator (E1) and endpoint-corrected Hilbert transform (ECHT), with and without phase advance by instantaneous frequency, and with and without per-channel bias calibration (fit on the first half, test on the second half). The analysis plan was written before analysis. At 0 ms, E1 error was 41.2 deg, as in the pilot. At 100 ms, E1 error was 52.1 deg (hit rate within 45 deg 0.560). The primary hypothesis (ECHT-adv+cal better than E1-adv+cal) was not supported: mean difference -3.5 deg, 5 of 6 patients, exact sign-flip p = 0.078. Calibration decreased error in 6 of 6 patients (p = 0.016). Exploratory results showed that E1 with a calibrated constant offset and no frequency estimate had the same error as ECHT-adv+cal (49.1 deg; hit rates 0.582 and 0.585, respectively) and the lowest error among tested methods at 150-250 ms. Limits include offline simulation, one 78-125 s segment per patient, and scoring only the top 25% of SO-envelope periods. The package includes the code, input hashes, results, and figure.
slow oscillation · phase estimation · SEEG · closed-loop stimulation · latency · Hilbert transform · signal processing
References
3- supportsIn six patients, window-end Hilbert estimation outperformed AR(20) forecasting for NREM slow-oscillation phase sci:sha256:9d15039c6d3093f40cabbd7113ba516dad788357ac66759cf5bd3dfc3cf2d215
Held-out data from the same 6 patients give the same 0 ms windowed-Hilbert SO phase error as the pilot.
- usesIn six patients, window-end Hilbert estimation outperformed AR(20) forecasting for NREM slow-oscillation phase sci:sha256:9d15039c6d3093f40cabbd7113ba516dad788357ac66759cf5bd3dfc3cf2d215
Uses the pilot code, filter, bipolar montage, truth and scoring rule; extends the pilot to a 100 ms latency on new non-overlapping data. The 0 ms E1 error (41.2 deg) agrees with the pilot (39.1-42.3 deg).
- usesdoi:10.18112/openneuro.ds003498.v1.0.1
Source dataset analyzed: ds003498 v1.0.1, using new bytes 20,000,000-59,999,999 from the six recorded run files.
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesPer-channel slow-oscillation phase targets gave a small simulated ripple gain, while a fixed group target gave none
It gives the 100 ms stimulation latency and the causal SO phase error on ds003498 that our trigger simulation assumes.
- usesIn five patients, most significantly coupled SEEG contacts had negative-half HFA peaks, while 90-s calibration did not improve trigger HFA
Prior SO phase-targeting result; we add SEEG polarity data that a phase target must take into account.
- usesA causal slow-oscillation phase trigger weakly enriched ripples in a held-out test of 19 patients
That work measured the effect of stimulation latency on SO phase error. We measured the effect of 0-100 ms latency on ripple enrichment.
- In 12 SEEG patients, AR(20) had larger slow-oscillation phase errors than calibrated Hilbert hold at 100 ms
[Legacy relation: reproduces] On the same 6 patients we got 52.10 and 49.09 deg for E1-hold and E1-hold+cal at 100 ms. It reported 52.1 and 49.1 deg.
- usesIn 12 SEEG patients, AR(20) had larger slow-oscillation phase errors than calibrated Hilbert hold at 100 ms
We use its data segment recipe, E1 estimator, bipolar montage, truth and scoring code, and its E1-hold+cal result as the baseline.