Cross-day SEEG decoding preserved AUC but worsened probability calibration in 10 subjects

We trained a shrinkage LDA decoder with Platt scaling on one SEEG session and tested it on another session. Data: OpenNeuro ds004752 (verbal Sternberg), archived late-maintenance log band power, label set size 8 vs 4, 10 subjects with sessions on 2 or more days, 334 train/test pairs. Main analysis uses a z-score for each session. On another day, AUC did not change (0.647 vs 0.641 within session, difference +0.006, 95% CI -0.035 to 0.047, p = 0.79). But calibration became worse. Excess ECE (ECE minus the ECE expected for perfectly calibrated outputs) increased from 0.004 to 0.053 (+0.049, CI 0.023 to 0.076, 9 of 10 subjects, exact sign-flip p = 0.0059). Brier score increased by 0.025 (p = 0.027). With raw features, the result was stronger: excess ECE +0.066 (10 of 10, p = 0.002), and the calibration intercept shifted by +0.42 (p = 0.031). Sessions on the same day were in between and not significantly different from within session. Thus, AUC alone does not show the cross-day problem for a BCI that uses decoder probabilities, for example in confidence gates or stopping rules. Limits: one task, one decoder, offline only, 10 subjects; the z-score uses unlabeled statistics of the full test session; we did not test recalibration. The package has code, input features with SHA-256 values, all rows, statistics, a figure and reproduction commands.

probability calibration · cross-session decoding · SEEG · shrinkage LDA · Platt scaling · expected calibration error · brain-computer interface