A 4-vs-8 SEEG decoder gave never-trained set-size-6 trials intermediate scores in 15 patients, but did not establish linear spacing

We tested whether a binary SEEG working-memory-load decoder gives a graded readout for a set size it did not see during training. We used archived log10 band-power features from the 3–6 s maintenance window for 15 patients performing a verbal Sternberg task in session 1 of OpenNeuro ds004752 version 1.0.1. A Ledoit–Wolf shrinkage LDA was trained on set sizes 4 and 8 only, using stratified 5-fold cross-validation with 20 repeats; set-size-6 trials were never used for training. The trained contrast had AUC 8 vs 4 = 0.701 (95% CI 0.607–0.796). Set size 6 scored above set size 4 (AUC 6 vs 4 = 0.614, 95% CI 0.538–0.690; 13/15 patients; one-sided sign-flip p = 0.0038) and below set size 8 (AUC 8 vs 6 = 0.597, 95% CI 0.526–0.667; 11/15; p = 0.0036). Correct-trial-only results were similar. The spacing difference was −0.017 (95% CI −0.113 to 0.079; two-sided p = 0.70), and the median position between class means was 0.54 among 10 patients with trained-contrast AUC > 0.6. This is not evidence of equivalence: the interval is wide and does not exclude a departure from linear spacing of about 0.1 AUC. Individual positions varied, with 3 of 10 patients reading set size 6 near set size 4. One-band differences were exploratory, not significant for spacing, and not corrected for multiple comparisons. The study used one session, one task, one dataset, offline analysis, and only 8–17 set-size-6 trials per patient; it did not test a closed-loop BCI.

SEEG · working memory · brain-computer interface · LDA · graded readout · OpenNeuro ds004752