In 11 SEEG patients, a fixed hand-knob distance prior improved held-out movement-versus-rest AUC over random channel selection

We tested a fixed anatomical rule for selecting SEEG channels without calibration labels. For each bipolar pair, we measured the distance from its midpoint in author fsaverage coordinates to the nearer of two population hand-knob points (−38, −22, 56) and (38, −22, 56), then selected the 8 nearest pairs (ANAT8). In 11 OpenNeuro ds006253 patients, the unchanged detector used log10 high-gamma activity, Ledoit-Wolf linear discriminant analysis, MOVE windows from 0 to 0.5 s after movement onset, REST windows from 0.5 s before stimulus onset, and a time-ordered split with the first 60% of trials for training and last 40% for testing. The preregistered primary result was mean held-out AUC 0.735 for ANAT8 versus 0.644 for the mean of 200 random 8-pair sets, a difference of 0.091; ANAT8 performed better in 9 of 11 patients (exact one-sided sign-flip p = 0.0088). Secondary tests found no reliable correlation between distance and training effect size (mean Spearman rho = −0.095, 8/11 negative, Holm p = 0.25; |d| versus distance rho = −0.036, Holm p = 0.50), and patient minimum distance did not predict all-pair AUC (rho = 0.15, p = 0.67). All-pair AUC was 0.824 (range 0.566–0.966), with per-patient permutation p < 0.05 in 11/11 and group sign-flip p = 0.00049. A label-based ORACLE8 selection had mean AUC 0.800; ANAT8 was 0.065 lower. No pair was nearer than 12 mm to the hand knob, and ANAT8 median distances were 20–43 mm, so the rule is a coarse prior for dorsal central cortex rather than a hand-knob sampler. This result applies to offline, task-locked windows, population coordinates, one mouse-movement task, and 11 patients; it supports a label-free first channel choice before calibration, not replacement of calibration. The planned addition of 11 new patients was not completed because the sub-07 download stalled and was stopped.

SEEG · channel selection · high-gamma activity · brain-computer interface · movement decoding · anatomical prior · held-out AUC · Ledoit-Wolf LDA · OpenNeuro ds006253