Causal linear repair reduced 50 ms offline 4–10 Hz phase forecast error with nominal 50 ms gaps every 250 ms in 36 selected SEEG patients

We tested repair of generated recording gaps in 36 selected SEEG patients from OpenNeuro ds004100 version 1.1.3. A fixed reader fitted on file A forecast phase in file B. Nominal 50 ms gaps repeated every 250 ms, with ten origin offsets. Linear repair joined the two gap endpoints only after the right endpoint was observed. Unfinished gaps remained held. Mean absolute error decreased from 61.44 to 60.13 degrees. The paired decrease was 1.32 degrees. Its descriptive 95% whole-patient bootstrap interval was 1.10 to 1.56 degrees. All 36 patient averages improved. Offline repair, which could use future endpoints, gave a larger decrease of 2.79 degrees. An extra label check after full-record mean removal gave a decrease of 1.28 degrees. Gaps were generated independently of voltage. No stimulation was delivered. The target used an offline 4–10 Hz phase reference at a nominal 50 ms horizon. This result does not establish benefit in a live interface. The package preserves 72 source clips, fixed models, main predictions, code, score tables, three figures, source hashes, and retrieval instructions. All 72 public source prefixes passed fresh checks. Models and the full analysis were reproduced.

phase forecasting · signal repair · stereo-EEG · brain-computer interface · recording gaps