SEEG power after key presses reduced error prediction loss beyond timing in separate blocks of four recordings

We tested rare key errors in OpenNeuro ds004473 version 1.0.2. An error was a valid key that differed from the arrow cue. Four recordings had errors in at least three task blocks. They supplied 1,057 trials and 28 errors. We tested each model on a separate task block. Inner block tests selected signal pairs and model settings. Adding 70 to 110 Hz SEEG power from 0.05 to 0.55 seconds after the key press reduced mean log loss beyond a timing model. Loss fell from 0.12864 to 0.09934 nats per trial, a 22.8% reduction. Total loss fell in all four recordings. In one recording, error-trial predictions became worse while agreement-trial predictions improved. Removing responses below 0.15 seconds preserved the mean gain. Before-key and before-fixation power had smaller mean gains. The two-sided patient sign-flip test gave p = 0.125. These results do not establish a population effect or live BCI correction. The package preserves code, features, source hashes, fitted models, every test probability, three figures, and reproduction instructions. Fresh public EDF ranges reproduced all 33 retained error trials and eight agreement trials exactly.

stereoelectroencephalography · rare error prediction · logistic regression · block validation · binary log loss · brain-computer interface