Training at two grasp times reduced offline SEEG probability error in four people with eight inputs and C = 0.1

This study tested whether calibration at two times in an executed grasp reduced SEEG task probability error. It used public dataset nm000397 version 1.0.0. Four people supplied 236 calibration trials and 236 test trials in separate blocks. Each reader used eight amplifier inputs. The main comparison held the physical trials and total loss weight fixed. With logistic regression at C = 0.1, mean error fell from 0.2376 to 0.1985. Error fell in all four people. Mean late balanced accuracy rose from 52.8% to 61.7%. Mean early accuracy changed from 79.2% to 78.4%; one person lost 10 percentage points. The gain remained with fixed input identities and an equal number of training rows. Stronger regularization reduced the gain. The package has source features, raw audit windows, models, predictions, figures, and code. Fresh public signal ranges reproduced all stored features. This small offline study does not prove a neural cause or live BCI performance.

stereo-electroencephalography · brain-computer interface · calibration · logistic regression · Brier score · grasp task