For an offline cued SEEG key-press LDA, margin flagged later-trial errors in 8 subjects; Mahalanobis novelty did not improve detection

We tested if a BCI decoder can flag its own errors. We used the archived 70-150 Hz and 13-30 Hz bipolar SEEG power of sci 04683a2a (OpenNeuro ds004473, 8 subjects, cued key press vs rest). A shrinkage LDA was trained on trials 1-100 and tested on trials 101+. The plan was written and hashed before the analysis. Primary results: a Mahalanobis novelty score did not flag decoder errors (median AUC 0.495, 4 of 8, Holm p = 0.28). A logistic combination of novelty and margin, fitted on trials 1-100, was not better than the margin alone (median AUC change -0.000, Holm p = 0.49). Secondary results: the margin |s| flagged errors in 8 of 8 subjects (median AUC 0.73). Rejection of the 20 % of windows with the smallest margin decreased the error rate of the kept windows by a median of 0.050 (8 of 8). The raw LDA posterior was overconfident (Platt slopes 0.09-0.40). Platt scaling fitted on out-of-fold trials 1-100 decreased the median ECE from 0.187 to 0.090 (7 of 8). Exploratory: novelty flagged errors only in the 3 subjects with decoder AUC > 0.9 (AUC 0.61-0.64). Limits: offline, cued fixed windows, 1 session, 8 subjects, 1 novelty definition.

SEEG · brain-computer interface · linear discriminant analysis · error detection · Mahalanobis distance · probability calibration