Under assumed timing, tuned SEEG forecasts had higher mean loss than task forecasts in 12 final numbered sessions

We tested forecasts of a wrong verbal memory response with public SEEG data from OpenNeuro ds004752 version 1.0.1. For each of 12 people, whole earlier sessions selected the ridge penalty. All scaling and selection of eight SEEG inputs stayed inside each training fold. The final numbered sessions had 561 trials and 49 wrong responses. Mean log loss was 0.2996 for tuned SEEG plus task data and 0.2924 for tuned task and past response data. Task minus SEEG loss was -0.0072 nats per trial; the descriptive 95% patient bootstrap interval was [-0.0281, 0.0102]. Tuning reduced mean SEEG loss by 0.0286 against fixed penalty 10. This reduction was absent with a stricter past-row quality rule. The study preserves all fold scores, model settings, forecasts, source records, figures, and code. The final outcomes were known from prior work. Forecast timing assumes that stored epochs start at fixation. Session date order was not verified. The result supports a test method for a brain computer interface; it does not establish a reliable error trigger or a stimulation benefit.

SEEG · verbal working memory · error forecasting · session transfer · ridge regression · patient bootstrap