Selection by matching effect signs showed no clear decrease in load probability error in 12 final SEEG sessions under assumed timing

We tested whether matching the sign of a load effect in two earlier SEEG sessions reduces probability error in a final session. We used OpenNeuro ds004752 version 1.0.1. The sample had 12 epilepsy patients, 36 sessions, 673 training trials, and 344 test trials. The targets were four and eight letters after a correct response. Each rule selected eight power features and used the same fixed logistic fit. Mean Brier error was 0.2587 for matching signs and 0.2531 for pooled selection. The paired difference was +0.0056, with a patient 95 percent interval of [-0.0050, +0.0180]. Six patients improved and six had higher error. The result gives no clear evidence of a decrease and does not prove equal performance. Controls used session mean removal, fixation power, and 100 random training partitions per patient. This is an exploratory test on a reused cohort. Phase labels assume fixation at the start of each stored epoch. The package includes features, source versions and hashes, code, fitted states, trial predictions, figures, and reproduction instructions.

stereo-EEG · working memory · brain-computer interface · feature selection · logistic regression · Brier score · session transfer · exploratory study