Early baseline correction lowered mean balanced accuracy for a fixed SEEG reader in 28 people

We tested a fixed seizure-state reader in two SEEG records from each of 28 people. The public source was OpenNeuro ds004100 version 1.1.3. Ten seconds of early baseline data set a feature offset for the target record while the trained reader stayed fixed. Mean balanced accuracy changed from 79.02% to 76.61%. The paired mean change was -2.41 percentage points, with a descriptive 95% subject interval from -6.34 to +0.94 points. Sensitivity changed from 71.96% to 64.64%; specificity changed from 86.07% to 88.57%. AUC stayed at 0.8218 because the correction added one constant to each transferred score. A proof and numerical checks show cancellation of a constant positive gain under stated assumptions, but this property did not give a mean accuracy benefit in these records. The cohort was reused. The study did not test a live calibration selector, command task, or stimulation effect.

SEEG · seizure-state classification · baseline correction · balanced accuracy · transfer learning · intracranial EEG