In 18 SEEG patients, training without primacy words lowered mid-list recall-decoder AUC (median difference -0.019, p = 0.027)

Question: a closed-loop memory BCI makes most decisions at list positions 4-12. Must its recall decoder be trained without the primacy words (positions 1-3)? Data: archived per-word band power of sci c4b8c884 (OpenNeuro ds004789 v3.1.0, FR1, session 0; 3-140 Hz, 0-1.6 s after word onset, all bipolar channels); MD5 sums agree. 18 of 19 patients passed the c4b8c884 inclusion rule. Method: L2 logistic regression (C = 7.2e-4, balanced), leave-one-list-out, AUC within serial position. Training sets: ALL positions, MID (4-12 only), PRIM (1-3 only), ALL75 (all positions, same size as MID). The plan was written before any fit. Primary result: on test words at positions 4-12, MID was lower than ALL (median AUC 0.597 vs 0.637; difference -0.019, 6/18 patients higher, exact Wilcoxon p = 0.027). A decoder trained only on primacy words decoded positions 4-12 above chance in 18/18 patients (median 0.606). The ALL decoder was not better on primacy words than on positions 4-12 (p_Holm = 0.58). MID vs the size-matched ALL75 had the same direction but was not significant (-0.009, p_Holm = 0.30). Thus we cannot separate the effect of more training words from more recalled training words (recall rate 0.34 at positions 1-3, 0.16 at 4-12). Conclusion: keep primacy words in the training set. Limits: offline, one session, one feature set, one fixed C.

SEEG · word recall · memory BCI · primacy effect · logistic regression · AUC · decoder training · OpenNeuro ds004789