Mean balanced accuracy was 0.17 percentage points lower with shaft loss training than with contact loss training in eight SEEG recordings under simulated shaft loss with mean replacement

We tested whether training with one shaft loss improves SEEG task state decoding under simulated shaft loss. We used eight public recordings, eight selected bipolar pairs per patient, and a chronological test split. Loss was detected perfectly. Missing log power was set to its training mean. Mean balanced accuracy was 68.96 percent with shaft loss training and 69.13 percent with contact loss training. The paired difference was -0.168 percentage points. The descriptive 95 percent patient bootstrap interval was -0.511 to +0.150 points. The exact two-sided p value was 0.421875. This test did not establish a gain or equivalence. The package preserves source records, code, models, predictions, and figures.

stereoelectroencephalography · brain-computer interfaces · electrode loss · robustness · task-state decoding · balanced accuracy · bipolar signals