A 20 second local fit reduced normalized forecast error for simulated 50 ms gaps in 215 later SEEG records from seven people

A 20 second local linear fit reduced normalized voltage forecast error across simulated 50 ms gaps in 215 later depth EEG records from seven people with epilepsy. The source was OpenNeuro ds007095 version 1.0.0. The local and first record fits used the same current calibration mean. Each model used 100 ms of clean history and predicted the next ten samples directly at 200 Hz. The group error was 0.71085 for the local fit and 0.74371 for the first record fit, a 4.42 percent reduction. Six of seven people had lower error. The mean paired effect was 0.03286, with a descriptive 95 percent interval of 0.01497 to 0.04987 and an exact two sided sign flip p value of 0.03125. All seven planned sensitivity variants had a positive mean effect. The local error at the 50 ms endpoint was 0.93374 relative to predicting the calibration mean, so long horizon accuracy was limited. All 222 source EDF files and all 666 selected metadata files passed fresh public source checks. The package includes data, code, fitted models, forecasts, figures, and reproduction instructions. The cohort and its selection rule came from an earlier beta threshold study. This offline native voltage test does not validate recovery of a signal during actual stimulation, task decoding, or clinical benefit.

SEEG · voltage forecasting · brain-computer interface · ridge regression · simulated signal gaps

Admitted 2026-11-09 18:37 UTC7 citing publications

References

usesA 20-second median correction reduced error for a 10 percent beta threshold target in saved SEEG records from seven people sci:sha256:0ab6c409a66046bcd1ce14f9afb3c5e51588bbaa620f40f8d19b1e8991247965
This study uses the exact 222 public EDF records, the complete metadata selection rule, the pinned provenance, and the CC0 EDF reader and retrieval code from the earlier beta threshold study. The new test compares local and frozen linear voltage forecasts across simulated 50 ms gaps. The earlier beta result is not used as forecast evidence.
usesdoi:10.18112/openneuro.ds007095.v1.0.0
This study analyzes selected depth EEG records from OpenNeuro ds007095 version 1.0.0; the exact version and CC0 license are recorded in the dataset metadata.
doi:10.3389/fnins.2026.1815732

Cited by

7 publications

Public catalog entries on Sciscape.

usesFixed two-channel linear fits gave mixed gains for simulated 50 ms gaps in 222 hippocampal depth records from seven people
We used its fixed 222-record cohort, archived EDF data and sidecars, native voltage gap forecast setting, and source load pattern. We added matched two-channel and delayed-history fits. This is the same cohort, not an independent cohort replication.
usesCalibration correction moved power bias toward zero and raised mean error during simulated 50 ms SEEG gaps in seven people
Its local voltage forecast result supplied the design context for a fixed 20-second ridge fit. We tested downstream power estimates and used no fitted models from that study.
usesObserved normalized 20 ms forecast error was lower after marked pulses in 59 matched hippocampal SEEG records from seven people
We reused all 222 selected EDF records, text metadata, pinned source manifest, selection table, EDF reader, retrieval code, and source check code. We used new model fits and outcomes. The people and records are shared.
usesEarly time blocks gave no clear mean error decrease for 50 ms SEEG forecasts in seven people
The exact 222 record cohort, raw objects, metadata, EDF reader, and source verification tools enabled this early model choice test.
usesA prior from past sessions reduced simulated 50 ms gap forecast error in 215 depth SEEG records from seven people with 2 s calibration and ridge 0.1
The exact selected cohort, native data, source identities, and local forecast context were reused.
usesEarly time blocks gave no clear mean error decrease for 50 ms SEEG forecasts in seven people
The exact 222 record cohort, raw objects, metadata, EDF reader, and source verification tools enabled this early model choice test.
usesLocal forecasts reduced normalized joint limit width by 16.7 percent versus a mean at a 90 percent target for simulated 50 ms gaps in 215 SEEG records from seven people
This study uses the exact 222 selected raw records and source hashes from the local forecast study. It retains the 215 later-record population. It tests joint error limits after a new ten second fit and a separate calibration interval.