More shafts drive most of the decoding gain from spread sEEG pairs in an eight-subject offline task

In an offline key-press-versus-rest decoding analysis of eight subjects from OpenNeuro ds004473 v1.0.2, we separated the effect of within-shaft pair spacing from the effect of sampling more shafts. We used archived high-gamma (70–150 Hz) and beta (13–30 Hz) log-power features and shrinkage LDA, with 100 random draws at budgets of 4, 8, and 16 bipolar pairs. A pre-result plan specified six primary tests, Holm correction, and support as Holm p < 0.05 with the same sign in at least 7 of 8 subjects. Spacing (scattered minus dense) increased mean AUC by 0.0076 at 4 pairs (8/8 subjects; Holm p = 0.047), but by only 0.0019 at 8 and 16 pairs (not supported). Sampling more shafts (spread minus scattered) increased AUC by 0.0323, 0.0522, and 0.0574 at 4, 8, and 16 pairs; this contrast met the pre-set rule only at 8 pairs (Holm p = 0.047; at 4 and 16 pairs Holm p = 0.070 and 0.0625). Spacing accounted for 19% of the spread-minus-dense gap at 4 pairs and about 3% at 8 and 16 pairs. A check restricted to scattered draws with no adjacent pairs found a +0.0146 AUC spacing contrast at 4 pairs (8/8 subjects, p = 0.0078); too few such draws were available at 8 pairs and none at 16. The replication contrast (spread minus dense) was +0.0399, +0.0540, and +0.0593 AUC at the three budgets. These offline results concern one task and eight subjects; shaft count is only a proxy for distance and brain region, and at 8 and 16 pairs scattered sets usually still contained adjacent pairs. For a fixed channel budget in this setting, sampling more shafts accounted for most of the decoding gain.

stereo-EEG · brain-computer interface · bipolar pairs · neural decoding · electrode placement · shrinkage LDA · AUC