A rolling 12-pair mean gave higher contrast prediction error than a fixed 12-pair mean at 3 to 15 ms in 14 of 20 early SEEG trains

We tested a past-only update for reading SEEG during electrical stimulation. We used 20 early train records from OpenNeuro ds005169 version 1.0.0. Both the fixed and rolling estimates used 12 completed pulse pairs. The next 12 pairs formed the test. The rolling mean gave a mean prediction gain of -0.58% at 3 to 15 ms, with a descriptive 95% person interval of -2.78 to +1.99%. Gain was positive in 6 of 20 people. Gain before stimulation was -3.75%. Their paired difference was 3.18 percentage points, with an interval of +0.89 to +5.80 points and an upper-tail pair-order probability of 0.0018. The positive paired difference came with losses in both conditions. It does not establish better correction during stimulation. Each update subtracts one contrast estimate from the first pulse and adds it to the second pulse. This keeps each pair mean at fixed detected edges. All 40 record and condition checks passed. All nine numerical files matched on replay. The package contains code, epochs, model estimates, source hashes, full scores, two figures, and retrieval steps. The test does not establish neural recovery, live edge detection, or brain interface benefit.

SEEG · electrical stimulation · pulse pairs · contrast prediction · rolling mean · signal processing · prediction error · brain-computer interface