Line removal prevented effects from added linear drift and increased clean error in 50 ms phase forecasts with 100 ms histories in 43 public SEEG records

We tested whether line removal can protect a small SEEG phase reader from added linear voltage drift. The data came from seven rest records in OpenNeuro ds006065 v1.0.0 and 36 clinical records in ds004100 v1.1.3. Each ridge reader used 25 block means from a past 100 ms or 500 ms history. It predicted an offline 4–10 Hz phase about 50 ms after the input cutoff. The reference used future voltage for evaluation. Line removal kept frozen forecasts fixed under additive linear drift across complete input windows. For 100 ms histories, it increased mean phase error on the original voltage by 6.86 degrees in the rest cohort and 7.93 degrees in the clinical cohort. The paired person bootstrap 95% intervals were [5.80, 8.10] and [6.79, 9.12] degrees. All 43 people had an increase. For 500 ms histories, the increases were 0.94 and 0.27 degrees. At three calibration slope SD values, line removal reduced error in both cohorts and both histories. This stress scale includes brain activity and does not estimate artifact rates. The result shows a clean accuracy cost of drift protection for these selected data and readers. It does not establish a stimulation benefit. The package preserves exact inputs, source records, a local plan, code, models, results, and three figures. New public source downloads checked all 43 input files. Independent fits checked all 688 models. All 279680 stream forecasts passed. A complete second analysis matched all 92 scientific output files.

SEEG · phase forecasting · linear drift · line removal · ridge regression · signal processing · brain-computer interfaces