Causal models predict marked SEEG ripple onsets only slightly better than chance in 20 patients

We tested whether causal signal features can predict that an author-marked ripple starts in the next 0–50 ms, which could help compensate for the approximately 37 ms latency of a causal ripple trigger. We analyzed the first 120 s of one interictal NREM SEEG run from each of 20 patients in OpenNeuro ds003498, using bipolar pairs with at least 10 marks. L2 logistic models used trailing ripple-band features (R), context features (C), or both (RC), with two-fold time-split validation. At 0–50 ms, RC had a median per-patient AUC of 0.554 (95% CI 0.549–0.581; 19/20 patients above 0.5). At 50% recall, median precision was 5.0% against a base rate of 4.1% (lift 1.20), so approximately 19 of 20 forecasts were false. Skill decreased with the horizon; at 100–200 ms, R was not different from chance (p = 0.10). Context features added a median AUC of only 0.006 at 0–50 ms (p = 0.13). The median score increased only in approximately the last 100 ms before onset. The prediction of R AUC above 0.75 was not supported. In this offline analysis of 120 s per patient, with linear single-pair features and author marks as labels, forecasts were not a practical way to compensate for trigger latency.

SEEG · ripple · forecasting · closed-loop stimulation · epilepsy · causal signal features · AUC · NREM sleep