A tuned nonlinear kernel increased mean relative squared error by 0.18% for future-only theta targets in seven SEEG rest records
We tested whether a nonlinear kernel improved a linear forecast of future SEEG voltage. The test used seven rest records from OpenNeuro ds006065 version 1.0.0 and four fixed pairs per person. Both models used 500 ms of past voltage and 56 s of early calibration. The target was the future-only part of a theta reference centered 50 ms ahead. The expanded rule included the linear rule. It selected a nonlinear kernel in 21 of 28 pairs. On later data, mean relative squared error increased by 0.18%. The descriptive person interval for error reduction was -1.28% to 0.54%. Linear skill versus a fixed mean target was 20.62%; expanded skill was 20.55%. The same method reduced error by 68.23% on a synthetic quadratic target. We include 19 spectrum controls, figures, checked source data, code, and saved models. This result applies to the tested model family and calibration budget. It does not prove that brain signals are linear or validate live brain stimulation.
SEEG · future-only forecasting · theta signal · kernel regression · Fourier phase controls
References
4- usesA 500 ms past voltage sum gave a 63.26 degree error for centered theta phase 50 ms ahead in seven SEEG rest records sci:sha256:678ac0e1b4e8d00b5e72eb22a8e9bbbfc9c4bd67c224a77925cb4cba677fc26f
Uses exact bipolar arrays, future-only target decomposition, source manifests, and source-check code. Reproduces its linear future-only MSE to numerical precision, then tests a nonlinear kernel under the same data split.
- usesLearning reduced mean 50 ms offline theta phase error by 11.25 degrees in spectrum controls from seven SEEG rest recordings sci:sha256:be46eec0bf0062fd4ea3fded570f754676cc2c3a9e4df6be657676ba2ae8d0c5
Uses its finding that spectrum controls explain much of learned phase performance to motivate full-pipeline Fourier phase controls for a nonlinear future-only forecast comparison.
- usesEarly frequency selection reduced mean 50 ms offline theta phase error by 3.55 degrees in seven SEEG rest recordings sci:sha256:5c3ae075960de0774e0fcc176ee07a5e3c9a353d86238f1e49023a23aa4d2bfa
Uses the fixed four-pair SEEG selection per person, made on the early theta power ratio. No new pair selection uses the later test interval.
- usesdoi:10.18112/openneuro.ds006065.v1.0.0
The analysis uses the seven rest records from OpenNeuro ds006065 version 1.0.0; the source record identifies the exact version and CC0 license.
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesEarly linear calibration with up to three extra SEEG pairs gave a 0.30% mean decrease in future theta forecast error in seven rest records
Supplies the exact seven-person bipolar voltage arrays, future-target implementation, source provenance, and own-pair baseline reproduced here.