Training for known 100 ms data age reduced mean phase error by 1.10 degrees compared with frequency advance for 2 s theta labels in two SEEG rest records

We tested whether training for known data age improves a theta phase forecast beyond frequency advance. We used two public rest records from OpenNeuro ds005670 version 1.0.0. The local plan was saved before the new score calculation. Four pairs per person were selected from early voltage. Models were fitted to early data. Later test decisions ran from 242 to 408 s. All methods used the same old 500 ms buffer. The action target was 50 ms after the decision. At 100 ms data age, frequency advance error was 75.89 degrees. Training for age error was 74.78 degrees. The mean gain across the two people was 1.10 degrees. The two record gains were -0.005 and 2.211 degrees. The package contains lossless input contacts, source versions and hashes, code, models, all forecasts, figures, and replay checks. The age was simulated and known exactly. The phase labels use future voltage. Two people do not support a population estimate. No stimulation effect or physical system timing was measured.

SEEG · phase forecasting · signal processing · brain-computer interfaces · data age