A 4–30 Hz decoder was not shown non-inferior to 4–150 Hz for verbal load in 15 SEEG patients

We compared offline shrinkage LDA decoders of verbal working-memory load (set size 8 vs 4) using archived 3–6 s log band power from 15 patients with stereoelectroencephalography (SEEG). Features covered 4–8, 8–13, 13–30, 30–70, and 70–150 Hz. With stratified five-fold cross-validation repeated 10 times and paired folds across feature sets, mean AUC was 0.662 for 4–30 Hz (LOW) and 0.708 for 4–150 Hz (ALL). The pre-registered non-inferiority test, with margin 0.05 AUC, did not show non-inferiority (mean LOW−ALL −0.045; one-sided p = 0.41; Holm-adjusted p = 0.66). The unplanned two-sided test found lower LOW than ALL (p = 0.046). The 30–150 Hz decoder had mean AUC 0.701; its difference from LOW was not significant (p = 0.33). Each single band decoded above chance after Holm correction, and 30–70 Hz alone had the highest mean AUC (0.682). Performance varied across patients. In this dataset, removing power above 30 Hz reduced mean AUC by about one fifth of ALL's above-chance AUC, while the median loss was 0.009. These offline results come from one session and one task in epilepsy patients with clinical electrode plans, with 9–29 trials per class, unequal feature counts between LOW and HIGH, and limited power to detect a 0.05 AUC difference. They do not establish performance for a real low-power implant.

SEEG · intracranial EEG · working memory · verbal Sternberg task · brain-computer interface · band power · shrinkage LDA · AUC · frequency bands