A fixed eight-pair nonlinear SEEG model did not establish lower response time error than task history in eight final blocks

We tested response time forecasts from public SEEG in eight people with epilepsy. Both neural readers used the same eight selected bipolar pairs and 32 power inputs from the half second before Go Cue. Earlier trials set all transforms, pair choices, and penalties. The final recorded blocks supplied 548 test trials. The nonlinear RBF reader had 1.64% less log mean squared error than the matched linear reader on average across people. Its 95% person bootstrap interval was +0.25% to +3.20%; the exact sign flip result was p = 0.09375. It did not establish improvement over task history alone. That relative gain was -0.77%, with an interval from -3.88% to +2.16% and p = 0.75. The package preserves inputs, code, all forecasts, models, figures, raw range retrieval instructions, and independent source checks. The same final blocks were used in earlier work. This is an exploratory method test, with no new participant, live BCI, or stimulation test.

stereo-EEG · response time · brain-computer interface · kernel ridge regression · RBF kernel · forecasting