Fixation hard negatives reduced false SEEG press responses in 8 patients, but press-versus-rest non-inferiority was not shown

In an offline analysis of eight epilepsy patients performing a cued key-press task, a standard shrinkage-LDA detector trained on press and rest windows scored the post-Fixation window above rest in all eight patients (mean AUC 0.607; p = 0.008; fixation score 0.30 of the press score). Training with post-Fixation windows as hard negatives increased held-out press-versus-fixation AUC from 0.778 to 0.818 (mean change +0.040; 8/8 patients; exact sign-flip p = 0.0078), with a similar increase using 70–150 Hz power alone (+0.048; p = 0.0078). A control that added the same number of rest windows had a smaller effect (hard-negative minus control +0.030; 7/8; p = 0.023), supporting a contribution from the fixation windows. Hard-negative training lowered the fixation score by 0.21 and reduced the fraction of fixation windows above the threshold for 90% press detection from 0.51 to 0.41. Press-versus-rest AUC changed by −0.008; the one-sided 95% lower bound was −0.0225, below the prespecified −0.02 margin, so non-inferiority was not shown. Fixation false alarms remained high in three patients with weak press read-out. The analysis used the first 100 trials from eight subjects in OpenNeuro ds004473 version 1.0.2, with five-fold cross-validation grouped by trial. Results are offline and limited to epilepsy patients, a cued task, one fixation window type, and 100 trials per subject; Go-cue windows were not tested as negatives.

SEEG · brain-computer interface · key-press detection · hard-negative training · shrinkage LDA · fixation false response · cross-validation · signal processing