Centered median templates raised mean site accuracy with one simulated wrong label per four fit pulses and lowered clean accuracy in five SEEG sessions at 20 to 100 ms

We tested whether a centered coordinate median reduces errors from wrong stimulation-site labels in a four-pulse fit. Public OpenNeuro ds004457 version 1.0.2 supplied five SEEG sessions, 183 sites, 732 fit pulses, and 732 separate test pulses. Each site had one wrong fit label in 64 draws per patient. Each draw reassigned recorded fit pulses and preserved all voltages. Both readers used the same mean-residual covariance fitted from the changed labels. At 20 to 100 ms, the median raised equal patient accuracy from 30.88% to 32.34%, a gain of 1.47 percentage points. Four patients gained and one lost. With clean labels, accuracy fell from 40.56% to 37.99%, a loss of 2.57 points. All five patients had a clean loss. The descriptive 95% patient bootstrap interval for the error gain was 0.25 to 3.05 points. The exact two-sided patient sign-flip value was 0.125. This small sample does not establish a population gain. The package provides a fixed calibration-label benchmark, three figures, voltage bins, all label maps and decisions, models, code, source hashes, and repeat instructions. Fresh raw extraction matched the input arrays. Direct analysis of the rebuilt inputs matched all 22 numerical files. An independent calculation matched all 566,568 saved decisions. These repeated decisions are not independent observations. The clinical cohort was reused. Only 36 test pulses occurred after all site fits. The study does not test a live BCI, real label faults, or a unique neural signal.

SEEG · brain-computer interface · stimulation responses · coordinate median · label errors · template classification · calibration · offline benchmark