Filters trained on other patients gave 56.29 degrees of mean 50 ms offline theta phase error in seven SEEG rest records

We tested shared theta phase filters in seven public SEEG rest records. Each test patient supplied four locally selected pairs. The pooled filter used phase targets, fit rows, setting checks, and thresholds from the other six patients only. It used 250 strictly past released samples for each forecast. At a 50 ms horizon, its mean absolute error was 56.29 degrees. A local filter with the same penalty gave 58.20 degrees. An added local tuning control gave 57.16 degrees. A pooled fit with only 560 rows gave 61.28 degrees, and a calibrated sinusoid gave 69.52 degrees. A fixed donor threshold accepted 25.06 percent of test times, with 40.33 degrees of error. The scoring reference was a centered two-second filter that uses future voltage. Contact selection still used early local data. These descriptive results test transfer of filter weights in reused data. They do not establish live signal performance or a stimulation benefit. The package preserves exact inputs, fresh public source checks, code, trained models, forecasts, three figures, and reproduction instructions.

theta phase forecasting · SEEG · intracranial EEG · ridge regression · cross-patient transfer · signal processing