In 10 SEEG subjects, 4–24 labelled trials did not restore cross-day calibration of a z-scored working-memory decoder

Question: can a short labelled block of k new-day trials restore the probability calibration that a SEEG working-memory decoder loses between days (sci 0126782b)? Data: archived late-maintenance band power of 10 subjects in OpenNeuro ds004752 (set size 8 vs 4), 208 other-day train/test session pairs. Decoder: shrinkage LDA with Platt scaling, as in sci 0126782b. We drew k = 4, 8, 16 or 24 labelled test-session trials (25 random draws and one chronological draw per pair) and evaluated on the remaining trials. Results with per-session z-scored features (main analysis): the pre-specified primary test failed. An intercept-only Platt refit with k = 8 did not decrease excess ECE (+0.007, 95% CI -0.012 to +0.025, 4 of 10 subjects lower, exact sign-flip p = 0.51), and it did not help at any k. A slope + intercept refit made calibration worse (excess ECE +0.116 at k = 4 and +0.064 at k = 8, 0 of 10 subjects lower, Holm p = 0.008), because the refit slope changed sign in 22-31% of draws. A new decoder trained on 8 new-day trials was worse than no recalibration (AUC 0.546 vs 0.650, excess ECE +0.092, Holm p = 0.006). With raw features, the intercept refit gave a small gain at k = 16-24 (excess ECE -0.022 to -0.024, 8 of 10 subjects lower, Holm p = 0.22). Limits: offline, 10 subjects, one task and one decoder, small evaluation sets (median 12-30 trials), z-score from the full test session, untuned L2 penalty of 1. The package contains code, input hashes, per-draw results, statistics, tables and the figure.

stereo-EEG · working memory · brain-computer interfaces · probability calibration · cross-day decoding · Platt scaling · shrinkage LDA