In offline FR1 decoding of 19 SEEG patients, random-word CV raised median recall AUC by 0.012 over list-grouped CV, but by 0.155 in one patient

We used the archived FR1 encoding band power (c4b8c884; 19 patients, 8 frequencies, all bipolar channels) and tested one ridge recall decoder with four CV schemes. The median held-out AUC was 0.666 with random words, 0.659 with random lists, 0.655 with contiguous list blocks and 0.637 with forward (past-lists-only) testing. Random-word CV was higher than random-list CV by a median of 0.012 (95% CI -0.001 to 0.035; 12 of 19; Wilcoxon p = 0.029, not corrected for 5 contrasts) and higher than forward testing by 0.018 (p = 0.023). Thus the inflation is small compared with the decoding effect. But the inflation was 0.155 in R1153T and 0.089 in R1070T. In 3 patients, list number alone gave AUC 0.66-0.70 under random-word CV. Over all patients, this list-number control did not significantly predict the inflation (rho 0.27, p = 0.26). Forward testing has fewer training words, so the forward result does not isolate a time-order effect. Use list-grouped or forward CV for closed-loop memory BCI estimates. Do a per-patient check with a list-number-only control.

SEEG · word recall · cross-validation · AUC · data leakage · memory decoding · ridge classifier · closed-loop brain-computer interface