In 18 patients, selecting 8 SEEG pairs did not improve recall decoding over random pairs

We tested how a word-recall decoder for a memory BCI changes when it uses fewer SEEG channels, and if selected channels are better than random channels. Data: OpenNeuro ds004789 v3.1.0 (FR1 free recall), session 0, bipolar pairs, 18 patients (88-173 pairs each). We used the archived per-word band power (3-140 Hz, 0-1.6 s after word onset) of sci c4b8c884. Decoder: L2 logistic regression (C = 7.2e-4), 5 folds of whole lists. In each training fold, we selected n = 1-64 pairs by univariate t or by all-pair decoder weight, and compared them with 20 random n-pair subsets. We wrote the plan before we ran any subset decoder. Primary test at 8 pairs: univariate selection was not better than random (median AUC 0.619 vs 0.606, higher in 12 of 18, exact Wilcoxon p = 0.28). Weight selection was also not better (0.610, Holm p = 0.50). Both 8-pair decoders were below the all-pair decoder (0.670; Holm p = 0.023 and 0.001). A half split (select on the first half of lists, test on the second half) gave the same result (p = 0.23). AUC increased approximately with log(pair count) for all rules. Selection helped a little only at 1-2 pairs (+0.03 AUC, descriptive). Top-8 sets from the two halves of a session overlapped little (Jaccard 0.067 vs 0.039 random). Thus recall information is distributed, and the number of pairs is more important than which pairs. Limits: offline, one session, two selection rules, 18 patients, clinical coverage. The package has numpy code, the plan, checksums, per-patient results and a figure.

SEEG · word recall · memory BCI · channel selection · AUC · intracranial EEG