In one session of 15 patients, SEEG load decoders showed no clear cross-period transfer, while pooled training showed no significant loss

We tested whether a decoder of verbal working-memory load trained during encoding also classified load during maintenance, and vice versa. We analyzed session 1, run 1 from all 15 subjects in the Zurich verbal Sternberg task (OpenNeuro ds004752), comparing set sizes 8 and 4 after removing trials marked Artifact==1. Five-band log power from SEEG channels was classified with shrinkage LDA and repeated stratified 5-fold cross-validation using shared folds. Within-period AUC was 0.605 for encoding (11/15 subjects above 0.5; Wilcoxon p=0.012) and 0.615 for maintenance (13/15; p=0.007). Cross-period AUC was 0.513 for encoding-trained, maintenance-tested (10/15; p=0.64) and 0.539 for maintenance-trained, encoding-tested (10/15; p=0.23). Maintenance-trained performance on maintenance exceeded encoding-trained performance there by 0.102 (p=0.048). A decoder trained on both periods had AUC 0.591 on encoding and 0.635 on maintenance; comparisons with within-period decoders found no significant differences (p=0.68 and p=0.42). One-second time generalization showed maintenance and probe blocks, while encoding-to-late-trial transfer was not evident; a late-encoding to early-maintenance effect was weak (AUC 0.54, uncorrected p=0.013). These results suggest period-dependent load patterns in these data and show no significant pooled-training loss, but do not establish that small cross-period transfer is absent. The analysis used one session and about 31 trials per subject, per-subject AUCs were noisy (about ±0.1), tests were not corrected for multiple comparisons, and only offline linear band-power decoding was tested. The complete code, results, figure, input-file URLs and hashes, and source sidecars are included; the 1.78 GB EDF files are not included.

working memory · SEEG · intracranial EEG · memory load · decoding · cross-period generalization · shrinkage LDA · time generalization