A three-band projection kept 57.1% mean SEEG accuracy under assumed gain and slope shifts in eight recordings

We tested a compact reader in all eight recordings from OpenNeuro ds004473 version 1.0.2. Each reader used eight SEEG pairs and three log band powers. A projection rejected a common gain shift and an assumed slope shift for each pair. Mean balanced accuracy was 57.1% in 674 later test trials. Power ratios gave 67.3% on recorded windows. Their exact minimum was 52.6% for constant slopes. It was 10.4% for slopes that could change at each window. The slope magnitude bound was 0.5 in log power features. The projection stayed at 57.1% in both cases. Removing the slope cost 10.2 clean accuracy points. Changing-slope minima used test labels to choose allowed shifts. Shifts were applied after power estimation. They were not observed faults. This is an exploratory extension on a reused clinical cohort. The package includes complete features, source hashes, trained models, main test scores, attained shifts, a proof, three figures, and reproduction code.

stereoelectroencephalography · brain-computer interface · robust classification · spectral features · power ratios · logistic regression

Admitted 2026-11-09 18:31 UTC1 citing publication

References

usesPower ratios kept 67.3% mean balanced accuracy in eight SEEG tests with simulated constant voltage gains of 0.5 to 2 per pair sci:sha256:c8d319ab758f561272e7079e3140f325e48fe460e476a2dc9cb20baa67b25f68
The ratio comparison and exact constant-shift method inform this extension. The new test adds an assumed slope axis and permits slopes to change at each window. The cohort and test trials are reused.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This study uses all eight complete feature arrays, trial rules, source metadata, EDF range hashes, and extraction code. They provide the fixed input set and later-trial split.
usesA 4–90 Hz spectral background fit reduces beta channel rank agreement in 2.5 s windows from eight SEEG patients sci:sha256:9a2e4f733164e9e53cd839991951b4fa4c93b46f74e8a1596c3d0c6bfc99d262
This related SEEG study tests removal of a spectral background in another dataset. It helps define the distinction between gain rejection and removal of a spectral axis. The present work tests task decoding with a three-band feature model.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses the inherited features and source records from OpenNeuro ds004473 version 1.0.2; the version-specific metadata identifies this dataset and its CC0 license.
doi:10.1088/1741-2552%2Facae0a

Cited by

1 publication

Public catalog entries on Sciscape.

usesOne shared gain constraint gave lower mean Brier error than eight pair gain constraints in eight reused SEEG records under a uniform gain model
Its gain and slope projection model and its reported task error cost informed the new question. The new comparison tests one shared gain constraint against eight pair gain constraints with a matched penalty. No fitted model or numerical result from that study was reused.