Thresholds matched to each simulated contact loss did not show lower mean background output in nine fixed SEEG readers

We tested thresholds after simulated contact loss in nine fixed SEEG readers. The public source was OpenNeuro ds003688 version 1.0.7. We tested all 109 single-contact losses in the selected readers. All pairs that used a lost contact had their four feature values replaced with training means. The test assumed a correct missing-contact flag. Thresholds matched to each loss changed the mean largest background rate from 14.35% to 15.44%. The paired change was +1.09 percentage points. Its 95% participant bootstrap interval was -2.06 to +4.14 points. This test did not show a mean background decrease or establish equal performance. A shared threshold prepared for all losses lowered the mean largest background rate to 7.07%. Mean heard-speech detection fell from 8.03% to 3.22%. The shared threshold bounds output on saved calibration windows. It does not bound future output. The package preserves all loss scores, frozen models, source features, three figures, raw audit samples, hashes, and repeat code. Independent score and raw feature checks passed. A separate process reproduced 13 numeric files exactly. This study reuses one clinical cohort and one film. It does not test intended commands or brain stimulation.

stereoelectroencephalography · brain-computer interface · contact loss · threshold calibration · speech detection · background output · neural signals · robustness