Mean offline task side decoding from SEEG before key presses exceeded a local label null in eight recordings

We tested whether stereoelectroencephalography (SEEG) band power before a key press identifies task side in later trials. We used all eight recordings in OpenNeuro ds004473 version 1.0.1 and 1,951 included trials. A fixed ridge decoder used three power bands from adjacent good contacts. It trained on the first 70% of trials and tested on the last 30%. Mean balanced accuracy was 56.0%. The descriptive 95% participant bootstrap interval was 50.4% to 63.2%. The local group permutation p-value was 0.003. This test assumes that labels can be exchanged within local trial groups, task condition, and train/test partition. Only participant 2 passed Holm correction. The result was uneven across participants. The before-response score did not clearly exceed the before-fixation score in a descriptive paired comparison. Random splits and after-response windows did not show a clear score increase. The source correctness flag rejected all 103 left-arrow trials with matching keys in participant 1. Response timing was known. The task cannot separate visual arrow processing from motor intent. The result does not establish continuous BCI control or an effect of brain stimulation.

stereoelectroencephalography · SEEG · brain-computer interface · task-side decoding · balanced accuracy · ridge regression · permutation test · OpenNeuro