Rejecting low-confidence SEEG trials increased balanced accuracy in 15 patients, while training thresholds undercovered

We evaluated a reject option for SEEG band-power decoders using archived features from 15 patients performing a verbal Sternberg task (OpenNeuro ds004752; one session per patient). With shrinkage LDA and repeated stratified five-fold cross-validation, retaining the 50% of test trials with the highest absolute decision scores increased balanced accuracy by 0.081 (95% CI 0.047 to 0.115; 12/15 patients; exact sign-flip p = 0.0011). Randomly retaining 50% produced no gain (+0.001; p = 0.89). The gain correlated with decoder AUC (Spearman 0.78); patients with AUC near 0.5 showed no gain. A threshold set from inner cross-validation scores increased balanced accuracy by 0.076 (95% CI 0.043 to 0.109; p = 0.0010), but realized coverage was 0.441 for a 0.50 target (difference -0.059; 14/15 patients below target; p = 0.0002). The likely explanation is that final-model test scores had less spread than inner-CV scores; calibration with the final model may address this bias. A second analysis of ds004789 free-recall features was incomplete after 16 of 19 patients because the shared computer became very slow. Descriptive results were mean AUC 0.60, balanced-accuracy gain +0.012 at 50% coverage, and realized coverage 0.43 for a 0.50 target (14/16 patients below target). Set A had 20-42 trials per patient and one session; its labels concern memory load, not motor or speech control. Set B has no completed pre-registered test. The publication includes the pre-analysis plan, code, results, checksums, and reproduction steps.

selective classification · abstention · SEEG · brain-computer interface · balanced accuracy · shrinkage LDA · cross-validation · coverage calibration · intracranial EEG · verbal working memory