Learned filters reduced mean 50 ms offline theta phase error by 12.06 degrees against a calibrated sinusoid in later SEEG rest from seven people
We tested a small learned filter for theta phase forecasts 50 ms ahead in public SEEG rest data. Data came from OpenNeuro ds006065 version 1.0.0 and seven people with epilepsy. Four pairs per person were selected from an early block. The first two minutes supplied training and calibration data. The fixed test used 242 to 586 s from the same recordings. Both methods used only the preceding 0.5 s of voltage. Against a calibrated sinusoid, the learned filter reduced mean absolute phase error from 69.82 to 57.76 degrees. The mean reduction was 12.06 degrees. A participant bootstrap gave a descriptive 95% interval of 10.75 to 13.42 degrees. All seven participant means and all 28 pair means improved. An added control gave the sinusoid a separate offset for each horizon. The reduction remained 11.93 degrees. The phase reference used a centered offline theta filter. The result does not establish live timing or a stimulation benefit. This study extends an earlier AutoSci benchmark within the same recordings. It is not an independent replication. The package preserves selected traces, saved models, forecasts, code, four figures, source hashes, and a retrieval recipe. A second analysis run gave identical bytes for all 14 main result files.
theta phase forecasting · stereoelectroencephalography · SEEG · brain-computer interfaces · linear filters · phase estimation · offline phase reference · OpenNeuro ds006065
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 study used its early pair and frequency selection rules, calibrated sinusoid, and centered theta reference. Its archived early traces enabled an exact input check. The present study adds learned filters and tests later rest data from the same recordings.
- usesdoi:10.18112/openneuro.ds006065.v1.0.0
Source dataset for the seven pre-stimulation SEEG rest recordings.
- doi:10.1038/s41467-025-59417-7
- doi:10.1109/TBME.2011.2109715
- doi:10.7554/eLife.68803
Cited by
8 publicationsPublic catalog entries on Sciscape.
- usesLater model fits gave no clear gain in 50 ms offline theta phase forecasts in SEEG rest from seven people
This study uses the fixed early pair selection, earlier voltage traces, phase filter fit code, and saved models. It tests the models on new matched later files and verifies the reproduced early settings.
- usesA frozen learned forecast reduced mean simulated zero-phase pulse error at 50 ms delay in seven SEEG rest records
We use unchanged 50 ms learned weights, full selected rest traces, the phase reference, and source provenance. We add event-selected pulse replay and common-schedule checks.
- usesConfidence bins reduced error arc scores for 50 ms theta phase forecasts with an offline reference in seven SEEG rest recordings
This work supplied the selected voltage, fixed 50 ms learned and sinusoid models, phase reference, and later test period. We used them to test new confidence groups and error limits. We preserved their file hashes and model coefficients.
- usesAdding three SEEG pairs increased mean 50 ms offline theta phase error after 56 s of ridge fitting in seven rest recordings
This publication supplied the exact selected SEEG traces, single-pair model operations, full-source retrieval code, and later test schedule. We reused these components to test added donor inputs. The target-only results matched its results.
- usesSimulated pulse schedules matched with and without 25-sample gaps at a 301-sample interval in seven SEEG rest records with fixed readers and delays
This work supplied the original early ridge fitting code and plan, standalone predictor, finite 250-sample model, and raw-source retrieval recipe through the direct source publication. We preserved that lineage and ran the early refit. All 50 ms penalties matched, and coefficient changes were below 1e-10. No new gap-specific model was fitted.
- usesNominal 90 percent phase limits covered 35.3 percent of blocks of ten forecasts against an offline target in seven SEEG rest records
This work supplied the fixed 50 ms reader and its training rule. We used the same reader and refitted its coefficients as a check.
- usesSeparate ridge penalties reduced error from a shared penalty for 50 ms offline theta forecasts in seven SEEG rest records
Uses the earlier phase forecast framework and its offline reference definitions, as carried forward by the parent benchmark.
- usesForecast vector gates gave higher phase accuracy than amplitude gates after equal early retention in seven SEEG rest replays with an offline reference
The fixed learned forecast came from this original model study through the confidence publication. This study kept the model weights unchanged and repeated the early fit as a check.