A linear readout agreed more closely with an offline theta reference than a local AR method at 20 ms in seven SEEG rest records

We tested theta phase forecasts for brain-computer interface timing in seven public SEEG rest records from OpenNeuro ds006065 version 1.0.0. We kept early learned linear coefficients from a cited AutoSci study. We added a local autoregressive method with 24 settings selected only on early data. Both methods used the same 0.5 seconds of available signal. At 20 ms on a new later 350 to 550 second interval, mean disagreement with the two-second offline reference was 52.12 degrees for the linear readout and 64.38 degrees for the AR method. The mean difference was 12.26 degrees. All seven records favored the linear readout. A descriptive person bootstrap interval was 11.15 to 13.34 degrees. The result persisted with a four-second reference. Linear disagreement exceeded 45 degrees at 43.95 percent of times. The same people were reused. Neural phase truth, live device timing, and stimulation benefit remain untested. The package preserves data, pinned source recipes, code, all forecasts, four figures, and exact reproduction checks.

theta phase forecasting · stereoelectroencephalography · SEEG · autoregressive models · linear readout · brain-computer interfaces · signal processing · phase estimation · OpenNeuro