In 15 SEEG patients, each power band decoded load above chance, but 4–30 Hz decoding was not shown non-inferior to 4–150 Hz at a 0.05 AUC margin

This study tested whether SEEG working-memory load decoding can use only 4–30 Hz power instead of five bands spanning 4–150 Hz. We analyzed archived band-power features for 15 patients performing a verbal Sternberg task in OpenNeuro ds004752, comparing set sizes 8 and 4. A patient-specific Ledoit–Wolf shrinkage LDA used stratified five-fold cross-validation repeated 20 times, with shared folds across feature sets. Mean AUC was 0.705 for 4–150 Hz, 0.662 for 4–30 Hz, and 0.694 for 30–150 Hz. The prespecified non-inferiority test for 4–30 Hz at a 0.05 AUC margin was not supported (p = 0.36); five of 15 patients lost more than 0.09 AUC. Each band alone decoded load above chance after Holm correction (mean AUC 0.632–0.675; corrected p ≤ 0.017), while removing any one band did not cause a significant loss. The direction of the load effect varied across patients and channels. These results show redundant load information across bands but do not establish that 4–30 Hz is sufficient. The study used one session, had 9–29 trials per class, fixed bands and a 3–6 s window, and did not recompute features from raw recordings. The 4–30 Hz features came from high-rate recordings, so performance of a real low-rate system may differ.

SEEG · working memory · band power · brain-computer interface · LDA · AUC · non-inferiority · OpenNeuro