A first-order model gave lower mean 1.5-second run forecast error in nine SEEG rest records at a 0.5 score threshold

We tested forecasts of the next three positive detector outputs in nine public SEEG rest records from OpenNeuro ds003688 version 1.0.7. The detector used separate half-second signal windows. We fit output probabilities to early rest and kept them fixed for later rest. At a score threshold of 0.5, mean squared probability error was 0.12626 under independence and 0.10980 under a first-order model. The mean gain was 0.01646. Its descriptive 95 percent person bootstrap interval was -0.00168 to 0.04475. A constant forecast from early run counts gave error 0.11242. An added error decomposition showed that 91 percent of the first-order gain came from a better mean predicted rate. At the fixed music threshold, later three-positive events occurred in only one person. The package contains all derived inputs, fixed neural models, code, forecasts, four figures, fresh raw audit bytes, source hashes, and repeat steps. Independent checks passed. An isolated run gave identical numerical files. The result concerns detector outputs during rest. It does not establish intended commands, clinical safety, or better stimulation.

SEEG · brain-computer interfaces · neural decoder · forecasting · Brier score · OpenNeuro