With mean replacement, one pair per shaft gave no clear gain in minimum SEEG balanced accuracy after simulated shaft loss in eight patients

We tested input selection for a small brain-computer interface decoder. Public SEEG data came from eight patients in OpenNeuro ds004473 version 1.0.2. Both rules selected eight bipolar pairs on early trials. One rule used signal rank alone. The other rule took one pair per shaft. We simulated loss of each eligible shaft in turn on later trials. Missing features used their training means, with perfect loss detection. Mean minimum balanced accuracy was 65.44% for rank selection. It was 65.58% for one pair per shaft. The paired gain was 0.14 points. The descriptive 95% patient bootstrap interval was -2.00 to 2.79 points. The exact two-sided sign-flip p value was 0.9453125. Four patients improved and four declined. The test did not establish a benefit or equal performance. A new full raw-range rebuild gave exact equality for all stored feature arrays in all eight patients. The package has three figures, all case predictions, fixed models, code, and source hashes. The scope is offline task-state decoding in one clinical session per patient. The test does not establish online control or hardware reliability.

stereoelectroencephalography · brain-computer interface · input selection · shaft loss · balanced accuracy · missing data