Causal 5 Hz low-pass filtering delays SEEG key-press decoding by 100 ms in three subjects, consistent with filter group delay

We compared causal and zero-phase pipelines for decoding cued left versus right key presses from stereo-EEG in four subjects (1, 2, 6, and 9) from OpenNeuro ds004473. Both pipelines used the same fourth-order Butterworth filters. In all three subjects with decodable low-frequency local motor potential (LMP) features, the causal AUC time course matched the zero-phase time course with a 100 ms delay, consistent with the 84–121 ms group delay of the 5 Hz low-pass filter. The causal pipeline reached the same peak AUC later. Across four subjects, mean AUC in the 300 ms before the press was 0.019 lower for the causal pipeline (zero-phase minus causal; exact one-sided sign-flip p = 0.0625, the smallest possible with four subjects). High-gamma differed by 0.001, but was near chance in three subjects. These results indicate that offline zero-phase analyses can show movement information about 100 ms earlier than an online pipeline can use it under this filter design and task. Limits include analysis of four of eight subjects because of a 45-minute run limit, one filter design, a single key-press task, and a 50-permutation null at one time point without correction across 49 time points; onset estimates are noisy and should not be interpreted. Key press does not establish motor-intention onset.

stereo-EEG · causal filtering · zero-phase filtering · brain-computer interface · key-press decoding · AUC · filter group delay