Early penalty selection reduced 300 ms cue SEEG forecast error compared with penalty 1 in eight reused patient tests

We tested whether early penalty selection could make 300 ms of cue SEEG power improve later key-delay forecasts. We reused OpenNeuro ds004473 version 1.0.2 features and known later tests from sci publication 0d5db2b1-6981-4f24-a4b3-12f8e53bc662. Eight patients supplied 1,002 early trials and 999 later trials. Three growing validation fits selected one of five penalties for each patient. All preprocessing used fit rows only. The selected cue model had mean absolute log error 0.18930; the task and response history model had error 0.19117. The paired gain was 0.00187 with a patient bootstrap interval from -0.00037 to 0.00468. Five patients had a gain. The selected model reduced error compared with fixed penalty 1, which had error 0.20332, in seven patients. Six patients chose the strongest tested penalty. A secondary early gate used power in four patients. The result shows sensitivity to the penalty but gives no clear added mean value over task history in this reused cohort. The test is conditional on eventual valid-key responders and does not test live control or stimulation. The package contains code, unchanged inputs, source versions and hashes, fitted states, forecasts, uncertainty draws, three figures, and reproduction instructions. Independent voltage and regression checks passed. A complete raw feature retrieval recipe is included; full EDF hashes were not checked.

stereo-electroencephalography · response delay · forecasting · ridge regression · brain-computer interface · model selection · patient bootstrap