Under assumed timing, eight SEEG features increased response time forecast error in 15 final sessions at ridge penalty 10

We tested whether SEEG power before a memory probe improved response time forecasts beyond task and past response inputs. We used public OpenNeuro ds004752 version 1.0.1: 15 patients, 68 sessions, 2,420 training trials, and 687 test trials. Each patient had a separate model trained on earlier numbered sessions and tested on the final numbered session. At fixed ridge penalty 10, eight training-selected maintenance power features increased mean squared log response time error from 0.12580 to 0.15252. The paired increase was 0.02671, with a 95% patient bootstrap interval of 0.01038 to 0.04914. Four of 15 patients had less error. More slope shrinkage reduced the increase; the interval included zero at penalty 1000. The test assumes fixation at epoch start and probe onset at second 6. It uses offline artifact flags and completed responses. This result applies to the tested model and split. The package provides figures, trial forecasts, fitted models, unchanged source features, pinned retrieval recipes, and separate numerical and fresh source checks. This is an exploratory test in a previously used cohort. It does not test stimulation.

SEEG · response time · forecasting · working memory · ridge regression · brain-computer interface · intracranial EEG