Offline SEEG readers trained on correct trials gave 48.6 percent balanced accuracy for inferred responses before movement on later errors in 15 MetaRDK records

We tested whether SEEG readers trained on correct trials could read inferred responses on later error trials. We used processed high gamma signals from public MetaRDK data, version 1.0.3. The primary test used 15 records and 356 later errors. Each person had a reader with eight channel means. Mean balanced accuracy before movement was 48.6 percent. Its descriptive 95 percent person interval was 42.7 to 54.6 percent. Accuracy after movement was 53.8 percent. The paired gain was 5.2 percentage points, with an interval of -2.5 to 15.3 points. Both primary Holm adjusted p values were 0.681. Accuracy before movement was also weak on later correct trials. The error result therefore cannot identify a special failure on errors. We assume that source correctness gives the binary response on retained trials with no recorded change of mind. The signals have offline baseline and smoothing steps, so this is not a test of online BCI performance. We replayed 5,364 earlier probabilities exactly and checked one fresh full source record. The package preserves compact data, source versions and hashes, retrieval code, fitted models, trial predictions, checks, and three figures.

SEEG · response decoding · error trials · brain-computer interface · balanced accuracy · MetaRDK