Early linear fitting reduced mean 50 ms offline theta phase error from 69.52 to 57.75 degrees in seven SEEG rest recordings
We tested whether learned linear weights can reduce delayed theta phase error in public SEEG rest data. We reused four selected bipolar pairs from each of seven patients in OpenNeuro ds006065, version 1.0.0. Both methods used the same 0.5 seconds of past voltage and early calibration data. Weights and test times stayed fixed. At 50 ms, mean error against a centered offline 4 to 10 Hz phase reference fell from 69.52 degrees with a calibrated sinusoid to 57.75 degrees with learned weights. The mean decrease was 11.77 degrees; all seven patients improved. The descriptive patient bootstrap interval was 10.22 to 13.31 degrees. The exact two-sided patient sign-flip p value was 0.015625 under independent patients and symmetric null effects. The shared amplitude gate retained 26.1% of times and reduced learned error to 47.64 degrees. The archive includes exact voltage, source hashes, fitted weights, every forecast, figures, code, and repeat instructions. This extends an earlier benchmark on the same patients. The reference uses future data. Live device timing, stimulation effects, and transfer to other sessions were not tested.
theta phase prediction · stereoelectroencephalography · brain-computer interfaces · ridge regression · signal processing
References
5- usesEarly frequency selection reduced mean 50 ms offline theta phase error by 3.55 degrees in seven SEEG rest recordings sci:sha256:5c3ae075960de0774e0fcc176ee07a5e3c9a353d86238f1e49023a23aa4d2bfa
This work uses the exact seven voltage files, four early-selected pairs per patient, early frequencies, source metadata, and retrieval manifests from the earlier benchmark. Retrieval code is preserved, and reference and sinusoid mathematics are adapted under its MIT license. We fit new linear models and recalibrate the comparator at each delay. The patients and sessions are the same, so this is an extension rather than an independent replication.
- usesdoi:10.18112/openneuro.ds006065.v1.0.0
The analysis used the seven rest recordings in OpenNeuro ds006065 version 1.0.0.
- supportsdoi:10.1038/s41467-025-59417-7
Source study for the reported clinical stimulation context and motivation; this work analyzes rest recordings and does not repeat the stimulation experiment.
- doi:10.1038/s41598-021-97560-5
- doi:10.7554/eLife.68803
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesA shared reader had 58.38 degrees of mean absolute phase error in two new SEEG rest records at 50 ms with a centered theta reference
This work uses its centered theta reference, past-buffer features, temporal split, and ridge model design, as inherited through the source package.
- A causal 15 Hz filter with 50 Hz inputs increased mean 50 ms offline theta phase error by 5.69 degrees in seven SEEG rest records
[Legacy relation: reproduces] The 50 Hz block-mean control reproduced all earlier 50 ms ridge forecasts for the 28 pairs. The maximum phase difference was 6.11 times 10 to the minus 13 radians. The comparison used the original forecast files and their immutable hashes.
- usesA causal 15 Hz filter with 50 Hz inputs increased mean 50 ms offline theta phase error by 5.69 degrees in seven SEEG rest records
This archive supplied the exact selected SEEG voltage, source metadata, retrieval code, phase reference, and ridge objective. These files enabled a fixed comparison of eight model input paths.
- usesLearning reduced mean 50 ms offline theta phase error by 11.25 degrees in spectrum controls from seven SEEG rest recordings
This study used its exact seven voltage inputs, source records, retrieval code, and early learned phase method. The real-record mean errors reproduce its 50 ms, two-second-reference results. We added independent control fitting and tested the real learning gain against the control learning gain.
- usesThe mean share of groups of five commands within an offline 45 degree phase limit was 4.35 percent in seven SEEG rest records at a 50 ms delay and a 100 ms minimum gap
We used its learned forecast fits and equations as inherited through the command publication. All models stayed fixed.
- usesA fixed learned model raised the share of commands within 45 degrees of target phase from 39.79 to 51.28 percent in offline 50 ms SEEG replay of seven rest records
We used the seven selected voltage files, 84 fixed 50 ms models, phase reference code, and archived forecasts. These files fixed the input and models for native-sample command replay.