Midpoint targets reduced the larger error across two theta phase references for 50 ms forecasts in seven SEEG rest records

We fitted past-only linear readers to the midpoint of two offline theta phase references. We used 28 selected bipolar pairs from seven public SEEG rest records in OpenNeuro ds006065 version 1.0.0. An early validation block selected the reference and penalty for the single-target comparator. At a 50 ms delay, mean larger error fell from 72.65 to 69.92 degrees. All seven person changes favored the midpoint target. Mean change was -2.72 degrees. A descriptive 95 percent person bootstrap interval was -3.35 to -2.16 degrees. The exact exploratory sign-flip value was 0.015625. Absolute error remained large. A wider peak band increased error for two people. The result applies to the specified filter labels. It does not establish true neural phase or stimulation benefit. The package contains exact voltage inputs, code, coefficients, time-level predictions, figures, and reproduction steps. Fresh full-source hashes and exact input rebuilds passed for all seven records.

theta phase · SEEG · phase forecasting · brain-computer interfaces · ridge regression · circular error · signal processing