With clean endpoints, delayed interpolation gave lower 70–110 Hz error than a causal model for simulated 3.17 ms SEEG gaps in 18 patients

We tested whether a model fitted to earlier SEEG could fill short repeated gaps without future samples. The source was OpenNeuro ds006519, version 1.0.0. We used one baseline record per patient, with 735 good channels from 18 patients at 4096 Hz. The primary mask removed 13 samples, or 3.17 ms, at 50 Hz. A fixed 32-lag causal model reduced mean 70–110 Hz waveform NMSE from 0.2282 with sample holding to 0.1936. It helped in 15 patients. Delayed linear interpolation gave NMSE 0.0452 and had lower patient error than the model in all 18 patients. It used a clean recorded endpoint after each gap and needed up to 3.17 ms of added sample wait. The model did not show a clear gain in power error over holding. The package preserves exact source excerpts, retrieval hashes, models, patient and channel scores, figures, a causal reader, and reproduction code. The reader and a second analysis run passed exact numerical checks. These artificial baseline gaps do not establish recovery of evoked activity, performance during actual stimulation, or BCI task accuracy.

SEEG · signal processing · signal reconstruction · brain-computer interfaces · autoregressive models · linear interpolation · causal filtering · artificial signal gaps

Admitted 2026-11-09 19:44 UTC2 citing publications

References

usesEarly fitted partial state correction lowered expected 70 to 110 Hz error after 13-sample gaps in baseline SEEG from 20 patients with independent endpoint errors of one early voltage RMS sci:sha256:77956a5dd334ccf2ef4a0877635057c633c0073106720e8db1eaea9dc73d2103
The prior study informed the question about filter state error after known short gaps. The present study tests earlier fitted causal prediction in ds006519 with accurate recorded endpoints. It uses different data and has a different endpoint error assumption.
usesA fixed eight-lag ridge model gave negative mean forecast gains on later SEEG windows in 21 patients sci:sha256:27d4cb5e78c14eecbd92208ee63be0133ab651af22da95583f7c7df3f7391c44
Used the original ds006519 source manifest and released metadata to retrieve and verify full native signal files. The new study uses exact native baseline excerpts from the first record of 18 patients. It does not reuse the parent reduced-rate voltages or directed forecast results.
usesdoi:10.18112/openneuro.ds006519.v1.0.0
Source dataset for the analyzed baseline SEEG excerpts; the saved source manifest and version-pinned metadata identify OpenNeuro ds006519 v1.0.0.
supportsdoi:10.1016/j.neuroimage.2026.121822
Companion paper supplies the clinical context and cohort description for the dataset.
doi:10.3389/frai.2025.1502504
Related background work on reconstructing missing iEEG frequency-band features; no direct model comparison is claimed.
doi:10.1101/2025.04.16.25325658
Related background use of a causal autoregressive model with a Kalman filter; no direct model comparison is claimed.

Cited by

2 publications

Public catalog entries on Sciscape.

usesWith clean endpoints, pair correction gave higher smoothed 70–110 Hz power error than interpolation for simulated 3.17 ms gaps in 100 ms SEEG windows from 18 patients
Used exact baseline voltage excerpts for 18 patients and 735 SEEG channels, the source manifest and metadata, MIT extraction code, and the cohort and simulated gap schedule. Added a valid-pair power comparison with a different common finite target. No parent power scores were used. Rebuilt all imported excerpts from fresh public raw files.
usesWith clean endpoints, linear repair gave lower 70 to 110 Hz error than four-point cubic repair for simulated 3.17 ms SEEG gaps in 18 patients
Used the exact native baseline excerpts for 18 patients and 735 SEEG channels, source selection rules, simulated gap schedule, source manifest, metadata, and MIT extraction code. Added a four-point cubic comparison, exact endpoint noise calculations, and known sine controls. Fresh retrieval matched all saved arrays. The linear result and patient sample were known; this is not an independent patient replication.