Memory load is decodable across SEEG regions, with no detected medial temporal advantage in 15 patients

We asked whether maintenance-period band power from stereo-electroencephalography (SEEG) contacts decodes verbal working-memory load in medial temporal (MTL), neocortical (NEO), and unlabeled (UNL, mostly white-matter) regions, and whether MTL outperforms an equal number of NEO contacts. We analyzed archived 3–6 s band-power features for 15 patients in OpenNeuro ds004752 (verbal Sternberg task, session 1, run 1), using set sizes 8 versus 4 and excluding set size 6. Shrinkage LDA with training-fold z-scoring and stratified 5-fold cross-validation repeated 10 times decoded load above chance in MTL (mean AUC 0.634; 12/14 patients above 0.5; exact sign-flip p = 0.0012), NEO (0.673; 11/15; p = 0.0040), UNL (0.668; 14/15; p = 0.0005), and all channels (0.700; 13/15; p = 0.0011). MTL did not outperform count-matched NEO subsets (MTL minus NEO AUC −0.022; 7/14 patients higher; p = 0.56). Mean absolute univariate effect size was slightly lower in MTL than NEO (difference −0.060, p = 0.036, uncorrected); no region-by-band result was significant after Holm correction. These results indicate that load information was present across regions in this task and that a decoder did not require hippocampal contacts in this dataset. The MTL-versus-NEO test had low power: it does not establish equality, and a difference of about 0.05 AUC cannot be excluded. Limits include 15 patients, one session, 9–29 trials per class, template-atlas labels, unequal channel counts except in the count-matched test, and use of IED-contaminated features. Individual region performance varied. All results, analysis code, source hashes, and reproduction instructions are preserved in the supporting files.

stereo-electroencephalography · working memory · memory load · brain-computer interface · decoding · medial temporal lobe · neocortex · band power · AUC