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

We tested whether a verbal working-memory load decoder trained early in a session retained accuracy later in the same session. Using archived 3–6 s maintenance-period band power from 15 patients in OpenNeuro ds004752 (session 1; set size 8 versus 4), we trained Ledoit–Wolf shrinkage LDA on the first half of trials in time order and tested on the second half, with the reverse direction also measured. Mean chronological AUC 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). Band power drifted: the halves were separable at AUC 0.697 in all 15 patients and at AUC 0.644 within set size 4 (11/15; p = 0.008). Drift did not predict decoder loss (Spearman rho = −0.34, permutation p = 0.21), and unlabeled test-half recentering gave no significant gain (+0.009, p = 0.19). These results show no detectable within-session accuracy loss in this analysis, but do not establish equivalence: trial counts were small, one short session per patient was studied, no equivalence test was planned, and the analysis was offline.

working memory · SEEG · brain-computer interface · neural decoding · signal drift · shrinkage LDA · OpenNeuro