Haufe patterns repeat better than raw decoder weights across list halves in 18 SEEG patients

In an offline analysis of free-recall (FR1) SEEG data from 18 patients in one session, we compared decoder maps fitted to the first and second halves of word lists. The data were 704–1384 log-power features per patient at eight frequencies. For shrinkage LDA, the Haufe pattern repeated better than the raw weight map: median Pearson r was 0.248 versus 0.033, and the Fisher-z difference favored the pattern in 17/18 patients (mean D = 0.224; exact one-sided sign-flip p = 7.6e-6). The result also held for 80+140 Hz, L2 logistic regression, and an odd/even list split. The pattern repeated slightly better than a Welch t map (p = 0.013). However, top-10% feature overlap was low (median Jaccard 0.065 for patterns), and weight stability had a stronger association with transfer AUC than pattern stability for LDA (Spearman rho 0.69 vs 0.34). Both maps exceeded a 200-draw label-shuffle null; the pattern null was about three times wider. A post-hoc comparison scaled by each map’s own null still favored the pattern, but with a smaller effect (15/18 patients, p = 0.0025). The secondary tests were not corrected for multiple comparisons. The study covers one task, one session, and 18 patients; it is offline, and its maps do not identify a reliable set of top contacts from half a session.

stereo-EEG · free recall · word recall · Haufe pattern · decoder weights · shrinkage LDA · feature stability · brain-computer interface