Four to eight screening lists improved 8-channel recall decoding over random channels in 18 SEEG patients, but performance remained below all channels

A memory BCI with fewer contacts must keep a small set of channels. We tested if a short labeled screening session can choose them. Input: archived per-word band power (8 frequencies, 0-1.6 s) of 18 SEEG patients from sci publication c4b8c884 (OpenNeuro ds004789, FR1, session 0). For each patient, an L2 logistic decoder (C = 7.2e-4) was trained on lists 1-8 and tested on lists 9-end. Z-scores used only lists 1-8. We chose the 8 bipolar channels with the largest summed squared Welch t on the first 1, 2, 4 or 8 lists. We compared them with 8 random channels (mean of 200 draws) and with all channels. The plan was written before the analysis. Result: median test AUC was 0.570 for random 8, 0.575, 0.599, 0.607 and 0.614 for 1, 2, 4 and 8 screening lists, and 0.621 for all channels. Channels chosen from 4 lists (48 words) and 8 lists (96 words) were better than random (+0.029, 14/18, Holm p = 0.0024; +0.035, 16/18, Holm p = 0.00015). 1 and 2 lists gave no significant gain (Holm p = 0.63 and 0.12). The 8-list choice was not shown non-inferior to all channels at a 0.02 AUC margin (median -0.019, p = 0.66). Exploratory: the chosen channels were not stable (median Jaccard 0.23 between the 4-list and 8-list choices) and almost did not overlap with the channels that were best on the test lists (0.07; random sets 0.04). But 8 channels chosen with test labels (leaky upper bound) gave 0.670. Thus a good small set exists, but 96 screening words do not find it reliably. Limits: one dataset, one session, offline AUC, one selection rule, one channel count, one fixed penalty. The package has the code, plan, input hashes, per-patient results and a figure.

stereo-EEG · word recall · channel selection · memory decoding · brain-computer interface · AUC · screening session