Haufe activation patterns repeated better than raw LDA weights across trial blocks in one-session SEEG data from 15 patients

In an independent SEEG dataset and task, Haufe activation patterns from the first 60% of trials repeated on the last 40% better than raw shrinkage-LDA weights. We analysed archived 70–150 Hz high-gamma features from 15 patients in OpenNeuro ds006253, comparing mouse movement with pre-stimulus rest within one offline session. Median map correlation across channels was 0.72 for patterns and 0.53 for weights; the mean Fisher-z difference was +0.335, positive in 14 of 15 patients (exact one-sided sign-flip p = 6.1×10⁻⁵), meeting the pre-registered replication criterion. Pattern stability did not exceed that of a univariate t map (mean difference 0.0001, p = 1.0). Top-10% channel overlap favored patterns by 0.107 Jaccard units (p = 0.021), while median pattern overlap remained 0.30. Map stability was associated with transfer AUC (Spearman rho 0.80 for weights and 0.69 for patterns). Combined descriptively with the earlier ds004473 result, D was positive in 20 of 23 patients. These findings support using activation patterns or univariate t maps rather than raw LDA weights for channel planning from clinical SEEG. The analysis was offline and covered one session per patient; movement and rest features came from different task phases, so some map differences may reflect task state. Only 70–150 Hz activity was analysed, and contact anatomy was not compared.

SEEG · brain-computer interface · linear discriminant analysis · Haufe activation pattern · decoder interpretability · high-gamma activity · movement decoding · map stability · OpenNeuro ds006253