Within-session band-power drift did not reduce working-memory load decoding in 15 SEEG patients

We tested whether a verbal working-memory load decoder trained early in a session retained accuracy later in that session. In 15 patients with stereo-electroencephalography (SEEG) from OpenNeuro ds004752, we decoded set size 8 versus 4 from 3–6 s maintenance-period band power using Ledoit–Wolf shrinkage linear discriminant analysis. Mean AUC for chronological first-half training and second-half testing (averaged with the reverse direction) was 0.679, compared with 0.690 for 200 random half-splits. The paired difference was −0.011 (median +0.023; 9/15 patients above zero; exact two-sided sign-flip p = 0.69). Features drifted: first versus second half was decodable at AUC 0.697 in all 15 patients, and within set-size-4 trials at AUC 0.644 in 11/15 (p = 0.008). Drift magnitude did not predict load-decoder loss (Spearman ρ = −0.34, permutation p = 0.21), and unlabeled test-half recentering gave a nonsignificant gain of +0.009 (p = 0.19). These results show no detectable group-level loss relative to random splits in this task and session; they do not establish equivalence. Each patient contributed 20–42 trials in one 272–400 s run, per-patient AUC was noisy, four patients lost 0.15–0.20 AUC, and the analysis was offline.

stereo-electroencephalography · working memory · brain-computer interface · signal drift · linear discriminant analysis · AUC · within-session decoding · OpenNeuro ds004752