Filters trained on other people gave 56.9 degrees mean 50 ms offline theta phase error in seven SEEG rest records

We tested whether theta phase filters can use training data from other people without local phase fitting. Public OpenNeuro ds006065 version 1.0.0 supplied seven people and 28 early-selected bipolar SEEG pairs. Each donor filter used early data from six people and excluded the target person. Later rest data supplied a 50 ms offline phase test. Donor mean absolute circular error was 56.931 degrees, compared with 56.928 degrees for local pair filters. The equal-person mean gap was +0.003 degrees. A descriptive resample interval was [-1.208, +1.253] degrees. Donor error was higher in four people and lower in three. This result does not establish equal performance. A local-person filter gave 55.769 degrees. One local offset reduced donor error to 55.106 degrees. Donor forecasts were within 45 degrees in 50.55 percent of cases. The offline target uses future voltage. Local unlabeled voltage was used to select pairs. Total calibration time was not shown to decrease. Source processing may use future voltage, and shared training folds limit inference. The package preserves voltage, source versions and hashes, code, and all results. It also supplies three figures and seven exported filters. Fresh source checks and a full numerical repeat passed. No live device or stimulation effect was tested.

stereoelectroencephalography · SEEG · theta phase · phase prediction · cross-person transfer · ridge regression · brain-computer interface · offline benchmark