Causal SEEG features weakly predict marked ripple onsets 20 ms ahead in 20 patients

This study tested whether causal SEEG features could predict author-marked ripple onsets before onset and reduce the approximately 37 ms detection lag reported in a prior analysis. We analyzed the first 120 s of run-01 for 20 patients from OpenNeuro ds003498: 481 bipolar pairs with at least 10 marked ripples, totaling 66,453 ripples. A per-pair L2 logistic regression used five causal features (80–250 Hz level and slope, 11–16 Hz level, and 0.5–2 Hz value and slope), with two time-split folds. At the preregistered 20 ms lead, median patient AUC against random non-ripple times was 0.565 (IQR 0.551–0.594); all 20 patient values exceeded 0.5 (one-sided Wilcoxon p = 9.5 × 10⁻⁷). Against same-ripple times 300–1000 ms before onset, median AUC was 0.558 (IQR 0.527–0.580; 17/18 patients above 0.5, p = 1.3 × 10⁻⁴). AUC decreased to 0.530 at a 100 ms lead. Ripple-band features alone gave AUC 0.567, while spindle and slow-oscillation features gave 0.529. At a 5% false-alarm rate, median onset hit rate was 8.9%. Thus, these five features and a linear decoder showed a weak pre-onset signal, insufficient to remove the detection lag. Limits include author marks as ground truth, only 120 s per patient, one decoder type, and instant-level AUC rather than a full event-level trigger simulation.

SEEG · ripple · closed-loop stimulation · causal features · logistic regression · intracranial EEG · onset prediction · OpenNeuro ds003498