Haufe activation patterns repeated better than LDA weights across early and late SEEG key-press trials in 6 of 8 subjects

We used the archived per-trial features of sci 04683a2a (OpenNeuro ds004473, 8 SEEG subjects, cued key press). We fit a shrinkage LDA (move vs rest) on trials 1-100 and a second LDA on trials 101 to the end. For each block we calculated the raw weights, the Haufe activation pattern (Cov(X) w) and a Welch t map. Then we calculated the Pearson r between the two blocks over the 100-112 bipolar pairs. The plan was written before the analysis. For 70-150 Hz features, the pattern repeated better than the weights (median r 0.67 vs 0.51; Fisher-z difference +0.44; 6 of 8 subjects; exact sign-flip p = 0.023). With 70-150 Hz + 13-30 Hz features the effect was larger (+0.57, 7 of 8, p = 0.008). Equal block sizes gave the same result. The top-10% channels by |pattern| overlapped more between blocks than the top-10% by |weight| (Jaccard +0.19, p = 0.047). A simple t map repeated as well as the pattern. Map stability followed decoder quality (Spearman rho 0.86-0.93 with transfer AUC). In the subject whose decoder was at chance (AUC 0.50), no map repeated. Thus, to find SEEG channels that carry key-press signal, use the activation pattern or t map, not the raw LDA weights. Limits: 8 subjects, one session, offline analysis, one decoder type; an odd/even split of trials 1-100 did not show a significant difference.

SEEG · brain-computer interface · decoder interpretation · Haufe activation pattern · linear discriminant analysis · map stability · intracranial EEG · key-press detection