Causal SEEG phase estimates had 49–64° error in two patients; amplitude gating and bias calibration reduced it to 29–42°

Question: How accurately can a closed-loop brain-computer interface estimate the phase of SEEG oscillations in real time, which method is best, and when is the estimate reliable? Data: OpenNeuro ds004752 (verbal Sternberg task), session 1, sub-01 and sub-02, bipolar adjacent-contact channels decimated to 250 Hz, 95 artifact-free trials. Method: For each channel, band (2-4, 4-8, 8-13 Hz) and evaluation time (2.0-6.4 s, 0.2 s steps), we compared 3 estimators that use only the past 1 s with a zero-phase FIR plus Hilbert phase on the full 8 s trial. The estimators are a causal Butterworth filter plus Hilbert (naive), the endpoint-corrected Hilbert transform (ecHT) and AR(30) forward prediction. Results: The best-method median errors are 48.9 deg (delta, ecHT), 54.5 deg (theta, ecHT) and 50.9 deg (alpha, AR). The naive method error is 60.6-64.3 deg. ecHT has a lower error than naive on 293 of 294 channel-band pairs. If only the top 25% of the causal amplitude is used and the per-channel bias (from odd trials) is removed on even trials, the best-method error decreases to 29.3, 35.4 and 32.5 deg. Limits: Only 2 of 15 subjects were processed in the time limit on a shared CPU, so the group tests have no power and the results are a within-subject benchmark. The reference phase is also an estimate, so the errors include reference error. The AR forward-prediction length was changed after a synthetic check.

SEEG · phase estimation · closed-loop stimulation · brain-computer interface · Hilbert transform · autoregressive prediction · intracranial EEG · OpenNeuro