In 7 patients, calibration ranking reached 90% of all-channel accuracy with 1–8 far SEEG channels, versus 8–32 for random sets

A bidirectional brain-computer interface (BCI) must verify which site it stimulated. We measured how many far SEEG channels (at least 15 mm from all stimulated contacts) are needed to identify one of 12 stimulation sites from a single 1–2 mA pulse, and whether calibration-based channel ranking reduces that number. Using 7 patients, 10 pulses per site, and an offline leave-one-pulse-out nearest-centroid classifier, random subsets needed 8–32 channels (median 16) to reach 90% of all-channel accuracy. Ranking channels by the training-fold F statistic reduced this to 1–8 channels (median 4). At k=4, mean accuracy was 0.410 ranked, 0.296 random, and 0.425 with all channels (chance 0.083). The primary ranked-minus-random contrast at k=8 was +0.062 (6/7 patients positive; exact sign-flip p=0.047). Secondary contrasts at k=1, 2, and 4 were +0.082, +0.130, and +0.114, respectively (Holm p=0.047 each). The pre-stimulus control stayed at chance. Results reproduce the archived all-channel accuracies. This is an offline, one-center result from 7 patients, 12 sites per patient, and 10 pulses per site; channel ranking used 119 calibration pulses, and the k grid was coarse. The data are from OpenNeuro ds005448 v1.0.0 (CC0).

SEEG · stereo-electroencephalography · brain-computer interface · single-pulse stimulation · site identification · channel selection · calibration · nearest centroid · epilepsy · mutual information