Three early sections of one film reduced mean later probability error in six SEEG recordings with eight selected pairs and equal sample counts

We tested how calibration section coverage affects a passive SEEG reader. We used six public recordings from OpenNeuro ds003688 version 1.0.7. Each reader used eight contact pairs chosen from its training data. One-section and three-section calibration each used 104 half-second windows. Three separate later film sections supplied every test. Mean Brier error decreased from 0.3211 to 0.2940 with three-section calibration. The paired decrease was 0.0271. Its 95% person bootstrap interval was 0.0126 to 0.0404. Five of six people had lower error. The exact two-sided sign-flip reference was 0.0625. Mean balanced accuracy increased from 57.28% to 58.84%, but the interval for this change included zero. The fixed-contact check had a smaller error decrease of 0.0042. Its interval also included zero. Both primary mean probability errors exceeded the 0.2500 constant-probability reference. These results apply to the tested contact choice and one film. Three-section sampling used equal retained signal time but more elapsed viewing time. The package includes three figures, features, raw source audit samples, all 768 models, all 239616 predictions, source hashes, code, and repeat instructions. Independent signal and score checks passed. A separate process reproduced all numeric outputs exactly. This study reuses six people from a cited package. It tests passive film labels and does not test intended commands or stimulation.

stereoelectroencephalography · SEEG · calibration · Brier score · brain-computer interface · speech and music · probability error · cross-validation