In 18 SEEG patients, recall decoding improved with each added bipolar channel; selection helped only at 1 to 2 channels

We measured how within-patient word-recall decoding changes with the number of SEEG bipolar channels (K = 1 to 32, and all). We used the per-word power features of publication c4b8c884 (OpenNeuro ds004789, FR1 session 0, 18 patients). The decoder was L2 logistic regression with leave-one-list-out tests. We compared random channel subsets (20 for each K) with channels selected in each training fold by a univariate t score. The median AUC increased at each step: 0.536 (1 random channel), 0.617 (8), 0.655 (32), and 0.669 (all). The primary test at K = 8 found no difference between selected and random channels (median +0.015, exact Wilcoxon p = 0.30). Selection was better only at K = 1 and 2 (median +0.057 and +0.055, Holm p = 0.039). All channels were better than random subsets at every K up to 32 (p <= 0.00025). Both methods needed 32 channels to get 90 % of the all-channel gain over chance. For this decoder and data, recall information is spread over many contacts, and a small set of selected contacts does not replace broad coverage. The package has the code (numpy only), input features, results, plan, and figure.

SEEG · word recall · memory decoding · bipolar channels · brain-computer interface · AUC · channel selection