In 10 SEEG subjects, pooled earlier-day training did not outperform the latest earlier-day session for new-day memory-load decoding

We compared training a memory-load decoder on all sessions from earlier days (POOL) with training on only the latest earlier-day session (LAST), using archived per-session band-power features from 10 subjects with SEEG data collected during a verbal Sternberg task. For the primary comparison, based on targets with at least two earlier sessions (31 targets), median AUC was 0.725 for POOL and 0.702 for LAST (median difference +0.018; 7/10 subjects; exact two-sided sign-flip p = 0.28). At equal trial count, POOL subsampled to LAST's class counts did not outperform LAST (median difference -0.023; 2/10; Holm-adjusted p = 0.52), while the exploratory POOL versus POOL_N comparison favored POOL in all 10 subjects (median difference +0.040; p = 0.002). Earlier-day training did not significantly reach same-day performance. Adding four labeled trials per class from the target day to POOL outperformed using those target-day trials alone (median AUC 0.712 versus 0.638; 9/10; Holm-adjusted p = 0.012). This comparison has the expected benefit of more training data; the planned analysis did not directly test POOL plus target-day trials against POOL alone. The results support a benefit from more earlier-day trials, but provide no evidence that more varied earlier sessions improve transfer. Results are limited by 10 subjects, 2–6 targets per subject, non-independent targets within subjects, confounding of day with time gap and clinical state, and use of one feature set, time window, and decoder without tuning.

brain-computer interface · SEEG · memory load · cross-day decoding · shrinkage LDA · band power · working memory · transfer learning