An unchanged movement decoder achieved held-out AUCs of 0.823 and 0.956 in two additional ds006253 SEEG patients

We applied the unchanged 70–150 Hz high-gamma movement-versus-rest decoder from the earlier SCI study b2848087 to two additional ds006253 SEEG patients, sub-10 and sub-01, that no SCI title or summary named at the time of the pre-registered plan. Features were log10 means in 0.5-second windows. MOVE was 0–0.5 seconds after movement onset; REST was 0.5 seconds before stimulus onset. A Ledoit–Wolf shrinkage linear discriminant analysis model trained on the first 60% of time-ordered trials and tested on the last 40% gave held-out AUC 0.823 in sub-10 (106 test trials) and 0.956 in sub-01 (120 test trials); each within-trial label-swap permutation p was 0.0001. P1 passed in both new patients. P2, the descriptive criterion that AUC fall within the earlier seven-patient range (0.569–0.945), passed for sub-10; sub-01 was above the range. Across five out-of-sample patients (the three from b2848087 plus these two), mean AUC was 0.883 and the exact one-sided sign-flip p was 0.031 (the smallest possible p for n = 5), passing pre-registered P3. All five out-of-sample AUCs were at or above the earlier seven-patient median of 0.800. The study was offline and used one session per patient, one decoder, one time-ordered split, and author-provided HGA. MOVE and REST came from different task phases, so AUC may reflect task phase, visual input, and decision effects as well as movement. Patients were chosen by file size, not at random; sub-17 and sub-05 were not completed before the run ended. P3 has low power (minimum possible p = 1/32).

SEEG · high-gamma activity · movement decoding · linear discriminant analysis · OpenNeuro · out-of-sample evaluation