Pre-confidence SEEG high-gamma activity did not reliably decode errors in five held-out patients

The base study found a weak trend in five patients from OpenNeuro ds006253: mean error-versus-correct AUC 0.55, sign-flip p = 0.094, with no patient-level permutation p < 0.05. We tested this result in five held-out patients using the unchanged code, features, classifier, and tests. The prespecified primary test did not pass: mean AUC was 0.528 (95% CI 0.447–0.608), one-sided sign-flip p = 0.281; 4/5 patients had AUC above 0.5 and 2/5 had permutation p < 0.05. The left/right choice control was also not significant (mean AUC 0.601, p = 0.0625). Pooled descriptively with the five base patients, mean error AUC was 0.539 (p = 0.0586), an effect that, if real, is small. HGA improved confidence-report decoding in the new group, but the groups disagreed and no multiple-test correction was applied, so no claim is made from this secondary result. The five-patient replication had low power: all five patients had to score above 0.5 for an exact sign-flip p < 0.05. Results are offline, within patient and session, use one prespecified time window and classifier, and use the authors’ HGA derivative without checking its preprocessing. Only five of 16 eligible new patients were analyzed because of download speed and time limits.

SEEG · high-gamma activity · error decoding · replication · confidence · brain-computer interface · OpenNeuro