In 12 SEEG patients, first-half calibration gave later-trial AUC 0.56, not above chance after Holm correction or significantly below random-split AUC 0.59

We tested whether a within-patient SEEG working-memory error predictor calibrated on early trials retained accuracy on later trials, as required for session-calibrated BCI use. We used archived per-channel band-power features from OpenNeuro ds006136 v1.0.1 (13 sessions, 12 patients) and a pre-registered per-channel shrinkage LDA ensemble. At the primary 3.5 s decision time, training on the first half and testing on the second gave median AUC 0.559; the one-sided test against chance had p = 0.039 and Holm-adjusted p = 0.077. Random 50/50 splits gave median AUC 0.594 (p = 0.0034). Chronological AUC was lower in 8 of 12 patients (median difference -0.039), but this difference was not significant (p = 0.23; Holm p = 0.23), and the study cannot exclude a loss of this size. At 6.5 s, a secondary, uncorrected analysis gave chronological median AUC 0.581 (p = 0.013) and random-split median AUC 0.611 (p = 0.0002); their difference was not significant (p = 0.34). Learning-curve comparisons found no significant differences. A larger change in error rate between session halves tended to coincide with a larger chronological loss, descriptively. This was an offline analysis of 12 patients, had low power for differences near 0.04 AUC, used trial index as a proxy for time order, and did not recompute features from raw SEEG.

stereo-EEG · working memory · brain-computer interfaces · error prediction · chronological validation · calibration · shrinkage LDA · AUC · OpenNeuro ds006136