A fixed learned update reduced 7 Hz filter phase error in nine SEEG rest records with simulated 100 ms gaps every 250 ms

Repeated signal gaps can leave errors in a filter state and its recent voltage buffer. We tested 36 selected bipolar pairs from nine public SEEG rest records. We kept the earlier fitted models fixed. The main test used 100 ms simulated gaps every 250 ms. Each causal reader kept its own estimated state after each gap. The learned reader filled older voltage gaps with the last received sample. Mean absolute phase error over eleven recovery offsets from 0 to 100 ms was 34.10 degrees for the learned update. Voltage hold had 42.24 degrees, and zero input had 37.50 degrees. The learned reader had lower mean error in all nine records than either comparator. Its mean error was 0.50 degrees above a diagnostic with exact state and history. All main-interval state transition norms were below 0.654 when voltage history stayed fixed. One matrix allowed transient growth at a 200 ms gap interval. Independent recursive replay, hidden-input checks, two full analysis runs, and fresh public source reconstruction passed. The package has data, models, code, source receipts, results, and three figures. These are exploratory results on reused selected rest records. They concern numerical filter phase, not biological phase, hardware recovery, decoder accuracy, or stimulation benefit.

SEEG · signal processing · filter phase · state estimation · recording gaps · ridge regression · neuroscience