Day-1 selection of four SEEG channels outperformed random channels on later days in 8 of 10 subjects, with weakly stable channel rankings

We tested whether a small set of bipolar SEEG channels selected on day 1 could decode verbal working-memory load on later days better than the same number of random channels. We used archived late-maintenance log band-power features for 10 subjects in OpenNeuro ds004752 v1.0.1, with 35 later-day sessions acquired 1–3 days after day 1. A shrinkage LDA trained on day-1 data was evaluated on each later session. At the prespecified primary budget of four channels, day-1-selected channels outperformed random channels in 8 of 10 subjects (mean AUC gain 0.07; exact one-sided sign-flip p = 0.008). Gains were larger at K = 1 (+0.12) and K = 2 (+0.095; 9 of 10 subjects for each; Holm-adjusted p < 0.01) and smaller at K = 8 (+0.03; 6 of 10; Holm-adjusted p = 0.047). For K = 1–4, selected-channel AUC was not significantly different from using all 30–59 channels; this is not evidence of equivalence. Day-1 and later-day channel rankings were weakly correlated (median subject-level Spearman rho = 0.15, positive in 10 of 10 subjects). This offline analysis covers one task, feature set, time window, and decoder, with small test sessions; day is confounded with time gap, medication, and clinical state.

SEEG · brain-computer interface · working memory · channel selection · shrinkage LDA · AUC · cross-day decoding · intracranial EEG