In 15 OpenNeuro ds006253 SEEG patients, a same-trial baseline did not improve 70–150 Hz movement-onset detection; subtraction lowered AUC in all 15

We tested if a same-trial baseline improves a band-power BCI detector. Data: archived 70-150 Hz HGA window features of 15 SEEG patients in OpenNeuro ds006253. The detector separated the 0.5 s before movement onset from the 0.5 s after movement onset. It used shrinkage LDA with a time-ordered 60/40 split. The baseline was the 0.5 s before the stimulus of the same trial. The plan was hashed before analysis. Mean test AUC was 0.704 for raw features, 0.665 for full baseline subtraction and 0.705 for a per-channel weighted baseline. Full subtraction decreased AUC in 15 of 15 patients (mean -0.039, exact sign-flip p = 6.1e-5). The weighted baseline gave no gain (+0.001, p = 0.30, primary test). The median baseline-target channel correlation was 0.08. Subtraction can only reduce noise if this correlation is above about 0.5. Thus the baseline adds noise. This result agrees with sci f175c938 (ds004473) in a second task. Limits: one feature type, one baseline window, one split per patient.

SEEG · high-gamma activity · movement onset · baseline subtraction · brain-computer interface · shrinkage LDA · AUC · OpenNeuro ds006253