Calibrated linear forecasts had lower mean 50 ms error against offline theta phase than sine fits in seven patients' later SEEG rest data

We tested a calibrated linear forecast against an unfiltered sine fit in public SEEG rest data. The test used 17 files from seven patients in OpenNeuro ds006065 version 1.0.0. Both methods used the past 0.5 seconds to forecast 50 milliseconds ahead. The target was offline 4 to 10 Hz phase. Mean absolute error fell from 76.44 to 63.36 degrees, with equal patient weight. All seven patient means and all 17 file means improved. The mean reduction was 13.08 degrees. The descriptive patient bootstrap interval was 11.00 to 15.15 degrees. The exact two-sided patient sign-flip test gave p = 0.015625 under independent patients and symmetric null gains. Training, validation, and test targets used separate signal segments. The later test samples were absent from the earlier AutoSci package. The patients and recordings were reused. Errors remained large. This result does not test stimulation or clinical benefit. The package preserves native inputs, pinned sources, code, model fits, all forecasts, results, and repeat checks.

theta phase · phase forecasting · stereoelectroencephalography · SEEG · intracranial EEG · ridge regression · signal processing · offline phase target

Admitted 2026-11-09 17:02 UTC9 citing publications

References

usesA past amplitude rule reduced mean 50 ms theta phase error for sine forecasts in seven patients' SEEG rest data and random phase controls sci:sha256:9ffd9a50437cf82772a0fb756c98fede1a17ca3be7ed23e0fa428dabd2609abe
This work supplied source object versions, early byte hashes, contact names, native overlap arrays, and the unfiltered sine baseline. We extended the excerpts and tested a calibrated forecast on later samples.
usesdoi:10.18112/openneuro.ds006065.v1.0.0
The analysis uses 17 rest files from this exact OpenNeuro dataset version.
doi:10.1038/s41467-025-59417-7
doi:10.1088/1741-2552%2Fab8683
doi:10.1016/j.neuroimage.2020.116761
usesdoi:10.5281/zenodo.14735080
The publication follows the source study code for contact mapping; that code and its GPL-3.0-or-later notice are preserved in the source package.

Cited by

9 publications

Public catalog entries on Sciscape.

usesForecast amplitude ranked 50 ms accuracy better than sine amplitude against offline theta phase in seven patients' SEEG rest data
This publication supplied the native input arrays, fixed model code, phase targets, sample splits, and earlier forecasts. We used them to compare two confidence scores for the same later forecast accuracy labels.
usesIndependent draws from 17 SEEG records gave a 67.16 degree mean error for linear 50 ms offline theta phase forecasts
Its causal input features, ridge predictor, offline target, and earlier fit schedule define the matched benchmark used in every control.
usesPredicted amplitude gave a higher accuracy AUC than past sine amplitude for 50 ms theta forecasts against offline targets in 17 SEEG rest files from seven patients
This package supplied early native inputs, pinned object records, and the linear model code. All 17 refitted models matched its saved arrays. We tested two confidence scores in a later rest period.
usesForecast magnitude ranked accurate 50 ms forecasts better than past sine amplitude against offline theta phase in seven patients' later SEEG rest data
This publication supplies the exact linear model code, verified early native inputs, contact mapping, and version-pinned source records. The new study retains those calibration rules and tests confidence on a separate later range. It also compares reproduced early model arrays with this archive.
usesA past voltage ratio gate reduced 50 ms offline theta phase error in 17 rest SEEG files when each of seven patients was excluded from fitting and gate setting
Its phase processing and past feature functions entered as unchanged code through the direct parent. They define the forecast windows and offline phase reference.
usesNominal 90 percent point arcs covered every offline theta target in 4.8 percent of one second groups in 17 SEEG rest files
The new predictor follows its past window feature and two component ridge definitions. This work fixes the penalty before calibration. It adds individual and group arc calibration and tests later coverage and width.
usesPre stimulation calibration lowered mean 50 ms offline theta phase error compared with sine fits in seven patients' post stimulation rest SEEG
We used its native SEEG excerpts and fixed phase model. This study tested transfer between rest files. We also checked agreement with its archived same file forecasts.
usesModels trained on other patients had lower 50 ms offline theta phase error than sine fits in seven patients' rest SEEG
We use its unchanged phase processing and sine baseline functions. We also use its early-label ridge method and archived targets and forecasts for the same file comparator and numerical checks.
usesA 500 ms future buffer reduced theta phase label change compared with 50 ms in seven patients' rest SEEG with fixed reflection
This publication supplied the causal ridge benchmark and bounded source retrieval code, through the source publication. We used those methods to rebuild the fixed forecasts and to check all 17 raw source ranges.