A 30 s rolling threshold decreased local acceptance rate error in seven later SEEG-labelled rest records at a 25% target

We tested a threshold based on the preceding 30 seconds of confidence scores. We used public ds006065 rest data from seven people with epilepsy and four selected SEEG-labelled pairs per person. Phase readers and the confidence model used fixed earlier weights. The nominal acceptance target was 25%. Mean absolute local rate error across eight 30-second blocks fell from 6.04 to 3.09 percentage points. All seven people had lower error. The paired reduction was 2.96 percentage points, with a 95% whole-person bootstrap interval of 1.95 to 4.49. Accepted phase error rose by 0.50 degrees on average. Its interval spanned zero and did not establish equal phase error. A periodic control had exact acceptance but higher phase error. This exploratory test used reused records and an offline filter phase label. It did not test stimulation benefit. The package preserves input data, fixed models, a causal stream gate, source receipts, code, tables, figures, and replay instructions.

stereoelectroencephalography · phase forecasting · confidence threshold · rolling quantile · acceptance rate · brain interface