In 15 SEEG patients, channel-band signs improved working-memory load decoding over one sign per band

We tested whether load-related band-power changes had a consistent direction across patients or channels, and compared decoders for verbal working-memory load. We used archived log10 Welch band power from bipolar SEEG channels in 15 patients performing a verbal Sternberg task (set size 8 vs 4; 3–6 s maintenance window; five bands from 4 to 150 Hz; 20–42 trials per patient). No band had a consistent sign across patients. High-gamma changes did not oppose theta or alpha changes over channels: correlations of channel-level t-values were positive and near the permutation null. Only high- and low-gamma effects were more similar than the null (r = 0.47 vs 0.35, 15/15 patients, p = 0.001). A decoder using one learned sign per channel-band had mean AUC 0.646, compared with 0.575 for one sign per band (difference 0.071, 14/15 patients, p = 0.003). Channel averages with shrinkage LDA had AUC 0.595; shrinkage LDA on all features had AUC 0.699. These results support learning channel-specific signs for this load-decoding task. The study used one session, task, time window, and five fixed bands, with few trials per patient. The analysis was offline, and p-values were not corrected across all tests; the main decoder contrast remained below 0.05 after Bonferroni correction across four decoder contrasts.

SEEG · working memory · verbal Sternberg task · band power · intracranial EEG · brain-computer interface · load decoding · shrinkage LDA · channel-band features