A calibrated causal filter gave lower 20 ms offline theta phase disagreement than a tuned harmonic model in seven SEEG records

We tested early calibration for later theta phase forecasts from public SEEG. We tested seven rest records from OpenNeuro ds006065, version 1.0.0. Each model used the last 250 samples. We fitted and selected models before 60 seconds. We then froze the models and tested all 2,000 clocks per person from 70 to 170 seconds. At a 20 ms forecast delay, a ridge filter reduced mean absolute disagreement with a 2 s offline phase reference from 71.35 degrees for a tuned harmonic model to 54.23 degrees. All seven people improved. The mean gain was 17.12 degrees; the descriptive person bootstrap interval was 15.64 to 18.91 degrees. Gains also held against a 4 s FIR reference and a Butterworth plus Hilbert reference. The package contains source contact excerpts, pinned retrieval records, model weights, tuning scores, all test outputs, four figures, and a streaming tool. This is an exploratory extension of a reused cohort. Offline reference phase is not biological truth. About 46% of learned forecasts still had error above 45 degrees. We did not test stimulation, behavior, or live hardware.

theta phase forecasting · stereoelectroencephalography · SEEG · ridge regression · causal filtering · signal processing