Calibration correction moved power bias toward zero and raised mean error during simulated 50 ms SEEG gaps in seven people

We tested a power correction during simulated 50 ms gaps in 222 saved depth SEEG records from seven people. Each record had two hippocampal channels at 200 Hz. A fixed ridge fit used the first 20 seconds. The correction added the propagated second moment of calibration forecast errors to causal beta power. It used no hidden test samples. Average power bias moved from -3.24% to +0.45%. Average normalized squared power error rose from 0.01904 to 0.02103, a 10.43% increase. Error decreased in three people and increased in four people. The forecast-minus-corrected gain was -0.001986. Its descriptive whole-person 95% bootstrap interval was [-0.005497, +0.000122]. The interval does not establish a population effect. Threshold sensitivity rose from 90.09% to 91.53%. False-positive rate rose from 0.573% to 0.735%. The correction can only raise fixed-threshold calls. Secondary curves showed that filter error grew after intact voltage returned. The formula is exact if the forecast has zero conditional mean error and the error matrix gives its conditional second moment. The fitted method need not meet these assumptions. The package includes source EDF files, fixed models, per-gap results, code, source hashes, retrieval commands, and three figures. Fresh source checks and an exact replay passed. This is an extension of an existing cohort. It does not test actual stimulation artifacts, behavior decoding, or clinical benefit.

signal processing · SEEG · brain-computer interface · causal power · forecast correction · calibration error · recording gaps