SEEG band power up to 1 s into the working-memory delay predicts later recall errors (median AUC 0.60, 12 subjects); the rest of the delay adds no significant accuracy after correction

We tested when single-trial SEEG predicts a later working-memory error. We used OpenNeuro ds006136 v1.0.1: 13 sessions from 12 subjects, a load-3 object memory task, and bipolar SEEG in temporal and frontal gyri. An error is a trial with fewer than 3 items correct. We wrote the plan before we computed features. We used the per-channel shrinkage LDA ensemble of 18254913 with 5-fold cross-validation. Primary test: a causal decoder that uses encoding plus the first 1 s of delay (decision time 3.5 s) gave a median AUC of 0.605. The AUC was more than 0.5 in 10 of 12 subjects (exact Wilcoxon, one-sided p = 0.0024). Encoding alone gave a median AUC of 0.581 (p = 0.0007). The full delay (6.5 s) gave 0.649. The gain from 3.5 s to 6.5 s was +0.030 (p = 0.034, Holm p = 0.068), which is not significant after correction. Each 0.5-s delay window alone gave a median AUC of 0.53 to 0.58, and the AUC did not increase over the delay (median Spearman rho 0.04, p = 0.52). Trial order did not explain the errors (median AUC 0.45). Conclusion: most error information is present at the end of encoding. A closed-loop memory aid can decide at 2.5 to 3.5 s with a loss of approximately 0.04 to 0.07 AUC. Absolute accuracy is low (AUC approximately 0.6), so single-trial use is limited. The package contains numpy-only code, the plan, per-channel features, per-subject results, SVG figures, and file MD5 values.

SEEG · working memory · memory error prediction · single-trial decoding · AUC · shrinkage LDA · causal decoder · OpenNeuro ds006136