Covariance gates increased mean rejection of faulty and recorded inputs under simulated contact gain 4 in eight offline SEEG tests

We tested two input gates on public SEEG data from eight people with epilepsy. The gates used 24 band-power features from eight bipolar pairs. Early trials set the pairs, decoder, and gates. Separate early trials set a nominal 5% rejection threshold. Later trials were held for testing. We changed one physical contact at a time before bipolar subtraction. Neither gate received the changed contact or the true class. At the primary voltage gain of 4, the covariance gate rejected 97.70% of faulty inputs. The diagonal gate rejected 95.76%. They also rejected 5.34% and 3.16% of recorded inputs. All rejection rates are equal person means. The primary difference in fault rejection minus recorded rejection was -0.24 percentage points. The descriptive 95% person bootstrap interval was -1.92 to 1.52 points. The exact person sign-flip probability was 0.796875. The primary test did not show a clear covariance advantage. Gains of 2, 0.25, and 0 gave different rejection tradeoffs. Many gain-loss inputs passed both gates. We preserve exact selected voltage samples, all scores, source checks, code, and three figures. The complete numeric repeat matched exactly. This is an offline test with artificial faults. It does not establish live BCI performance.

brain-computer interface · stereoelectroencephalography · input rejection · covariance gate · fault simulation · abstention · contact gain