An early linear reader reduced unseen input squared error by a mean 0.89% for a centered 7 Hz, 1 s target with a 50 ms center horizon in 250 hippocampal SEEG records

We tested whether an early linear reader could predict unseen filter input beyond the exact known input control. The test used 250 public hippocampal SEEG records from seven people. The centered target used a 7 Hz carrier and a 1 s Hann window. Its center was 50 ms after the forecast cutoff. Its support extended 550 ms after the cutoff. Each reader used early data and predicted later data. The mean relative reduction in unseen input squared error was 0.888%. The descriptive 95% person bootstrap interval was 0.441% to 1.377%. All seven person means were positive. Mean phase error fell from 54.347 to 54.158 degrees. The same procedure lost accuracy on independent Gaussian noise. This exploratory test reused a cohort. It measures a defined voltage target and does not establish biological phase accuracy or stimulation benefit. The package preserves native data, source versions and hashes, code, fitted models, predictions, figures, and repeat checks.

SEEG · phase prediction · signal processing · linear prediction · ridge regression · forecasting · epilepsy