SEEG working-memory load AUC showed moderate patient stability across sessions in 10 patients

We assessed the session-to-session reliability of a shrinkage-LDA decoder of working-memory load in 56 SEEG sessions from 10 patients with at least two sessions. The decoder classified Sternberg set size 8 versus 4 from late-maintenance band power. Session AUC had an ICC(1) of 0.52 (95% CI 0.11 to 0.67; permutation p = 1e-4), with a true between-patient AUC standard deviation of 0.113. Within-patient AUC spread exceeded Hanley–McNeil trial-noise estimates (Cochran Q = 80.7, df = 46, p = 0.001), but did not exceed the cross-validated AUC noise estimated under shuffled labels (Q = 48.1, df = 46, p = 0.39). Same-day session pairs did not agree better than other-day pairs (mean |dAUC| 0.111 versus 0.127; p = 0.50), and first-session AUC did not significantly predict the mean AUC on other days (Spearman rho = 0.35, p = 0.16, n = 10). Spearman–Brown estimated 3.7 sessions, about 140 trials, for reliability 0.8. These findings support using multiple sessions for patient or electrode-site screening in this setting. The estimate is limited by 10 patients, one task, one feature set, one decoder, and the uncertainty of the cross-validation noise model; the ICC confidence interval is wide and has a lower bound near zero.

SEEG · working memory · decoding · AUC · reliability · intraclass correlation · shrinkage LDA · brain-computer interface · session variability