Eight selected bipolar channels retained all-channel word-recall decoding accuracy in 10 SEEG patients

We tested how many stereo-electroencephalography (SEEG) channels a patient-specific decoder needs to predict later word recall during encoding, and whether a fixed anatomical region is sufficient. We used archived per-word band-power features from 10 patients performing the FR1 delayed free-recall task in one session each. In leave-one-list-out evaluation, the eight channels selected from training lists had median AUC 0.668, compared with 0.670 for all 55–125 channels (median difference −0.012; exact two-sided Wilcoxon p = 0.92; 4/10 patients better). This result does not establish equivalence. Eight selected channels outperformed eight random channels in all 10 patients (median difference +0.048; one-sided p = 0.001). Within serial position, AUC was 0.613 for selected channels and 0.610 for all channels; selected channels again outperformed random channels in all 10 patients (p = 0.001). One selected channel performed below the all-channel decoder (median difference −0.055; p = 0.049). In seven patients with atlas labels, four-pair decoders limited to medial temporal, lateral temporal, or frontal sites did not differ consistently (median AUC 0.52, 0.59, and 0.57; all pairwise p > 0.6), and the best region varied among patients. The results support using patient-specific calibration to select a small channel set in this dataset. Limits include 10 patients, one session and task, low power for equality claims, a single selection rule, and leave-one-list-out training that can use lists recorded after the test list.

stereo-electroencephalography · word recall · memory decoding · brain-computer interface · channel selection · band power · leave-one-list-out · anatomical region