Full contact delay training reduced later worst Brier error beyond positive weight scaling in eight SEEG state readers at a simulated delay of four samples

We tested whether full training with simulated contact delays gave lower error than positive scaling of the clean feature weights. The restricted fit had a free intercept and used the same delayed training inputs, weights, scaler, and L2 penalty as the full fit. We used OpenNeuro ds004473 v1.0.2: eight people, 1,332 early training trials, and 674 later test trials. At a delay of four source samples, about 4.002 ms, mean worst contact Brier error was 0.3255 for the restricted fit and 0.2493 for the full fit. Full minus Restricted was -0.0762, with a descriptive 95% person interval [-0.1082, -0.0431]. All eight differences were negative. Mean contact AUC fell from 0.7601 to 0.7389. Clean Brier error rose from 0.1869 to 0.1964. The full mean worst error was close to 0.25, the error of a constant probability of 0.5. The data and later trials have extensive prior use. This exploratory result applies to the tested offline task states and fault set. It does not establish live BCI benefit or a stimulation result. The package contains code, features, source range records, model weights, predictions, tables, figures, and replay instructions.

stereo-EEG · SEEG · contact delay · Brier error · logistic regression · brain-computer interface · fault tolerance · probability calibration