Whole-shaft training showed no clear gain over matched pair omission in lowest balanced accuracy after known shaft loss with mean replacement on later trials from eight SEEG recordings

Whole-shaft training showed no clear gain over matched pair omission under known simulated shaft loss. We used eight recordings from public SEEG dataset ds004473, version 1.0.2. Each reader used eight pairs and three power bands. Earlier trials trained the reader. Later trials tested it. Training means replaced missing log-power features. The primary training omission probability was 0.5. The exact control matched the omitted-pair count distribution and each pair's marginal omission rate. The mean lowest balanced accuracy was 65.24% for shaft training and 65.29% for matched training. The paired change was -0.05 percentage points. The person-bootstrap 95% interval was -0.62 to 0.47 points. The exact two-sided sign-flip p value was 0.9375. Mean intact accuracy was 70.89% and 70.66%, respectively. No reader gave a constant class in any tested shaft-loss case. A full raw rebuild matched all 3,133,395 stored power values and all 2,032 used range hashes. The package preserves data, code, masks, readers, predictions, figures, provenance, and instructions. This reused cohort gives exploratory evidence for offline reading. It does not test stimulation or a live interface.

stereoelectroencephalography · SEEG · brain-computer interfaces · signal loss · logistic regression · robustness · balanced accuracy