In 15 SEEG patients, 4–30 Hz power was not shown non-inferior to 4–150 Hz for decoding verbal working-memory load

We compared shrinkage LDA decoding of verbal working-memory load (set size 8 vs 4) from archived maintenance-period (3–6 s) log band power in 15 patients with stereo-EEG (SEEG). The features were five bipolar-channel bands (4–8, 8–13, 13–30, 30–70, and 70–150 Hz) from session 1 of OpenNeuro ds004752, version 1.0.1. Stratified five-fold cross-validation was repeated 10 times with shared folds across feature sets. Mean AUC was 0.708 for all bands (4–150 Hz), 0.662 for low bands (4–30 Hz), and 0.701 for high bands (30–150 Hz). The planned one-sided non-inferiority test of low versus all bands, with a 0.05 AUC margin, did not show non-inferiority (mean difference −0.045; p = 0.41; Holm-adjusted p = 0.66). High minus low bands had mean difference +0.039 (p = 0.33; Holm-adjusted p = 0.66). In an unplanned two-sided test, low-band AUC was lower than all-band AUC (p = 0.046); high bands were not different from all bands (difference −0.007, p = 0.79). Each single band decoded above chance after Holm correction; 30–70 Hz had the highest single-band mean AUC (0.682). Differences varied across patients. These offline results come from one session and one task, with 9–29 trials per class, clinical electrode plans, fixed band edges, unequal band counts, and limited power to detect a true 0.05 AUC difference. They do not establish performance for a live low-power implant.

stereo-EEG · working memory · neural decoding · band power · AUC · shrinkage LDA · intracranial EEG · frequency bands