Reaction-time deconvolution links pre-movement SEEG detector scores to movement preparation in 15 patients

We used trial-to-trial reaction-time jitter to separate movement-locked and stimulus-locked parts of a held-out 70–150 Hz shrinkage-LDA movement-detector score in archived SEEG features from 15 patients. The movement-locked kernel was above rest from 0.4 to 0 s before movement onset in all 15 patients (median 0.58; one-sided exact sign-flip Holm p = 6.1e-5), reaching a median 0.74 of its value at 0–0.4 s after onset. At a fixed time to movement, the pre-registered stimulus-locked 0–0.6 s window was not above zero (median −0.01; 6/14 positive; p = 0.80). The group-median stimulus kernel was descriptively +0.2 to +0.3 at 0.4–1.2 s; this later interval was not tested. The mean cross-validated MOVE-versus-REST AUC was 0.85. Exploratory checks with a 4 s stimulus kernel and without the weak sub-20 detector did not change the movement-locked result. These findings suggest that much of the early score can reflect movement preparation beginning up to about 1 s before onset, rather than a stimulus-locked response in the first 0.6 s. They do not show that no stimulus response exists. This offline analysis used one session and one task per patient, archived 0.2 s window features rather than continuous data, and assumes additive movement- and stimulus-locked components that do not vary with reaction time; stimulus effects were constrained to zero before stimulus onset and from 2 s afterward. Decision activity that ends at movement is classified as movement-locked. Sub-14 had no eligible H1 estimate because reaction times left fewer than 10 observations in required stimulus columns; sub-20 had a weak detector (AUC 0.54).

SEEG · movement preparation · brain-computer interface · reaction time · deconvolution · high gamma activity · shrinkage LDA · stimulus-locked · movement-locked