SEEG did not consistently outperform scalp EEG at decoding working-memory load in an 11-patient pilot

This pilot compared stereo-electroencephalography (SEEG) and simultaneous scalp EEG for decoding verbal working-memory load during the maintenance period of a modified Sternberg task. We used OpenNeuro ds004752 v1.0.1 (15 epilepsy patients; 3 s maintenance window), streaming only set-size 4 and 8 trial data for the analyzed sessions. Log band power was decoded with ridge-regularized linear discriminant analysis and repeated stratified 5-fold cross-validation. Results are reported for sub-01 to sub-11, with 17–68 correct artifact-free trials per patient. Median AUC was 0.75 for EEG and 0.71 for SEEG across all five bands. AUC exceeded the 95th percentile of a 40-permutation null in 5/11 patients for EEG and 6/11 for SEEG; SEEG all-band AUC exceeded EEG in 7/11. No modality exceeded the null in four patients. High-gamma-only SEEG had a median AUC of 0.62. These results do not show a consistent SEEG advantage. Per-patient trial counts were low, null thresholds ranged from 0.60 to 0.87, and no paired group test or multiple-comparison correction was performed. The analysis covered only some sessions and did not include the planned channel-count and region analyses.

working memory · SEEG · scalp EEG · brain-computer interfaces · AUC · permutation test · epilepsy