A leave-one-subject-out SEEG decoder predicts working-memory errors above chance in 10 of 12 patients (median AUC 0.60)

Can a working-memory error decoder transfer to a new patient without labeled calibration data? This study tested leave-one-subject-out (LOSO) prediction using OpenNeuro ds006136 v1.0.1, with 13 subjects and 14 sessions performing a load-3 object working-memory task. The analysis used 10 log-power features from theta, alpha, beta, low-gamma, and high-gamma bands during encoding and delay. Features were z-scored within each session without labels, averaged across the available channels, and z-scored again. For each held-out subject, a shrinkage linear discriminant classifier was trained on all other subjects; the two sessions from MU2 were held out together. The test set contained 13 sessions from 12 subjects. The decoder was above chance in 10 of 12 subjects (median AUC 0.602; one-sided Wilcoxon p = 0.0034), and 8 of 13 sessions had permutation p < 0.05. Performance was similar to within-session five-fold cross-validation with the same features (median AUC 0.593; paired p = 0.47), and slightly below the prior per-channel ensemble (median AUC 0.642; paired p = 0.064). Across all 13 training folds, higher low-gamma and lower theta and alpha weights pointed toward errors. Trial order did not explain the result. The z-scoring used all unlabeled trials in each test session, including scored trials; a causal calibration-block version was not tested. The evidence comes from 12 subjects, one dataset, and one task. Muscle activity may contribute to gamma power, and further artifact control was not done. The secondary fixed-direction scores are not independent of the prior work on this dataset. The archived code, features, results, provenance records, and figure are included.

working memory · SEEG · intracranial EEG · error prediction · cross-subject decoding · leave-one-subject-out · shrinkage LDA · band power · brain-computer interface · human neuroscience