Huber regression reduced 50 ms theta phase forecast errors under simulated 10 ms, ±1 mV faults in 36 SEEG records

We tested two phase readers on 36 public SEEG records from OpenNeuro ds004100 version 1.1.3. Each reader used the past 500 ms to predict a theta phase 50 ms later. We added separate positive and negative 1 mV faults to the final 10 ms of each input window. The offline phase reference always used the original voltage. The primary score used the larger error across the two signs at each cutoff. The mean error was 129.8 degrees for weighted least squares and 79.6 degrees for Huber regression. The paired difference was -50.2 degrees, with a descriptive person bootstrap interval of [-52.50, -47.99] degrees. All 36 person means improved. On original voltage, the difference was -0.83 degrees. With a 50 ms fault, the Huber error rose to 124.8 degrees. The package contains exact source samples, code, figures, source checks, and an exact numerical repeat. This input fault test does not measure true neural phase or a stimulation benefit.

Huber regression · phase prediction · stereoelectroencephalography · SEEG · theta phase · robust estimation · signal processing · input faults