A pre-pulse spectral-state model did not choose larger CCEP N1 pulses better than a sham control

In 8 SEEG patients, we tested whether a ridge model using 15 pre-pulse phase and log-magnitude features at 4, 6, 10, 15, and 20 Hz could select larger single-trial CCEP N1 responses (10–50 ms). The archived input contained 1,198 response-site units at least 10 mm from the stimulated pair, with 12,379 real unit-trials and 12,379 sham windows shifted by −0.25 s. In leave-one-patient-out validation, the top-half versus bottom-half gating gain was 0.091 SD for real pulses and 0.042 SD for sham windows. Their difference was 0.048 SD (95% bootstrap CI −0.068 to 0.157; exact one-sided sign-flip p = 0.223; 5 of 8 patients). The pre-registered rule required p < 0.05 and mean real gain ≥0.10 SD; it therefore found no usable gating gain. The one-sided 95% upper bound for the real-minus-sham gain was 0.141 SD. These results apply to the archived single-trial table and its 3-cycle Morlet features, response-aligned phase, template-projection N1, and sham-window procedure. Limits include 8 patients, clinical 1–2 mA pulses, response-site features, a linear model, and patients with epilepsy.

closed-loop stimulation · CCEP · N1 response · SEEG · brain-computer interface · spectral state · ridge regression · sham control · epilepsy