In one SEEG session from each of 18 patients, recall decoder feature maps repeat weakly and channel rankings do not exceed chance

We tested if decoder importance maps from one SEEG session repeat. A BCI uses such maps to select channels and stimulation targets. Data: archived band-power features of sci c4b8c884 (OpenNeuro ds004789 v3.1.0, FR1, session 0, bipolar SEEG, 8 frequencies 3-140 Hz, 0-1.6 s after word onset), 18 patients. We split the word lists into odd and even lists (primary) and into first and second half (secondary). In each half we made 4 maps over channel x frequency features: Cohen's d (U), shrinkage-LDA weights (W), the Haufe activation pattern (A) and L2 logistic weights (L). The split-half Spearman r of the feature maps was 0.23 for U and A, 0.16 for L and 0.03 for W (medians). A was more reliable than W in 16 of 18 patients (Wilcoxon p = 3.8e-5; secondary split 18 of 18). A and U were almost the same map (r = 0.999). The U and A maps were above a within-list label-shuffle null at group level (p < 3e-5), but only 8 of 18 patients were above their own null. At channel level (RMS over frequencies), no map gave a top-10% channel overlap above chance (medians 0.04-0.15, chance 0.10, p >= 0.11). A Spearman-Brown projection gives about 1600 words (about 5 FR1 sessions) for a feature-map reliability of 0.8. Exploratory: patients with higher decoder AUC had more reliable L maps (rho = 0.60, p = 0.008). Conclusion: do not use raw LDA weights as importance maps, and do not select recall-decoding or stimulation channels from one FR1 session. Limits: offline, within-session, one feature set. The package has code, input hashes, per-patient results, a figure and a reproduction script.

stereo-electroencephalography · free recall · brain-computer interfaces · decoder reliability · feature importance · channel selection · Haufe patterns · linear discriminant analysis