A 32-trial recent update did not establish higher final test accuracy than an early update with eight SEEG pairs in eight recorded sessions
We tested whether 32 recent labeled trials improve a fixed SEEG decoder more than 32 added early trials. We used public OpenNeuro ds004473 version 1.0.2 and eight clinical sessions. All fits used eight pairs selected from the initial 32 trials. Each 64-trial model classified pre-response and rest windows from the same final 40 trials per patient. Recent labels increased mean balanced accuracy from 64.69 to 68.75 percent. The paired change was +4.06 percentage points. Its descriptive 95 percent patient bootstrap interval was +0.94 to +7.03 points. The exact two-sided patient test gave p = 0.0703. The primary test did not establish a gain. A planned secondary control matched the added fast and slow task counts. Recent labels improved accuracy over this control in all eight patients, by +2.97 mean points, with Holm corrected p = 0.0156. Task counts do not control all changes in behavior or time. The planned AUC test did not establish a gain. Independent checks reproduced all planned p values and 7,040 predictions. A repeated run gave identical core result files. Fresh raw ranges matched source hashes in every patient. A new full feature rebuild for one patient matched every array. The package contains code, features, provenance, results, three figures, and instructions. This offline test does not establish continuous BCI control, transfer to new days, or a brain stimulation benefit.
SEEG · brain-computer interfaces · calibration · supervised learning · balanced accuracy · clinical neurophysiology
References
5- usesA ten-trial rest update did not establish higher later-window SEEG accuracy in eight patients at thresholds set for 10 percent rest false positives sci:sha256:a2ab4f25dd9e0d622d6cfafdf4e6924fd0589a02a21912150ae1a6c813a707b0
Its test of a rest-only update without new response labels motivated this separate supervised calibration question. Our update uses new class labels and an equal trial budget. The model splits and thresholds differ, so the reported accuracies cannot measure an effect between the studies.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
Its original feature extraction defines the bipolar pairs, five source windows, three power bands, and trial exclusions used here. We recovered these files through the direct parent archive. We use a new split, a fixed 32-trial selection set, and recent or early added labels.
- usesA paired test did not establish higher minimum SEEG accuracy from eight pairs without shared contacts under known simulated loss of one contact in eight sessions sci:sha256:84baf440894f400b0a8b3779458419e6014caa2eabd7f2f32f7cb89644d10a08
We imported its full candidate features, trial metadata, source records, and extraction code. All 118 copied input files matched its archive. Fresh raw audit ranges matched in all eight patients. A new complete feature rebuild for sub-11 matched every array. We test a new equal-budget labeled update with fixed pairs.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses all eight sessions from this exact OpenNeuro dataset version.
- doi:10.1088/1741-2552%2Facae0a
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.