A 100 ms voltage history gave 1.16 degrees less mean offline theta phase error than 500 ms in seven SEEG rest records at 25 inputs and a 50 ms forecast

A 100 ms voltage history gave 1.16 degrees less mean offline theta phase error than 500 ms in seven public SEEG rest records. Each history gave 25 block means. Separate ridge models forecast phase 50 ms after a past-only input cutoff. We used 28 fixed adjacent pairs from OpenNeuro ds006065 version 1.0.0. Mean absolute error was 56.59 degrees with 100 ms and 57.75 degrees with 500 ms. Six of seven people had less error with 100 ms. The descriptive 95 percent person bootstrap interval for the short-minus-long difference was -1.81 to -0.25 degrees. The exact two-sided sign-flip value was 0.078125 under independent people and symmetric null differences. Point and native-sample controls had intervals that include zero. History length also changes temporal averaging. The short method used 80 percent fewer persistent float64 voltage buffer bytes: 400 instead of 2000 per pair. Both models had 25 inputs and 50 output weights. These counts do not measure device energy or total memory. Only 51.42 percent of short-history forecasts were within 45 degrees of the reference. The package includes selected voltage arrays, all 840 models, all forecasts, three figures, a past-sample reader, source hashes, and reproduction code. An independent solver checked every model. A second run matched all numerical arrays and result tables exactly. Fresh complete downloads checked all seven raw files and exactly reproduced all 3,360,000 stored voltage values. This is an exploratory extension on a reused cohort. The phase reference uses future samples. New sessions, hardware timing, behavior, and stimulation effects were not tested.

theta phase forecasting · SEEG · intracranial EEG · voltage history · ridge regression · phase error · brain-computer interface · temporal averaging