A fixed ripple threshold lost 0.035 mean F1 on the last interval in 12 SEEG patients; losses were large in 2

We tested whether a causal ripple-trigger threshold calibrated on the first NREM interval stayed accurate on the last interval in 12 patients with SEEG recordings from OpenNeuro ds003498. We analyzed the first 60 s of the first and last runs, using author ripple marks as the reference, across 252 bipolar pairs. On the second 30 s of the last run, mean patient F1 was 0.381 for a frozen threshold, 0.393 for label-free re-normalisation, 0.416 for labelled recalibration, and 0.428 for a within-run reference. The frozen-minus-recalibrated mean difference was -0.035 (median -0.019; 95% CI -0.070 to 0.014; exact two-sided sign-flip p = 0.081). The 90% CI of the mean (-0.065 to -0.007) was not within the ±0.05 equivalence margin, so the data did not show that the loss was smaller than 0.05. Seven patients lost F1 and five gained. Losses were large in two patients, where a lower envelope median made the frozen threshold too high and recall fell. Label-free re-normalisation recovered an average 0.012 F1 versus the frozen threshold, with no consistent direction. These results apply to short segments from 12 of 20 patients with smaller montages. The reference was offline author marks, not ground truth, and no stimulation was delivered.

ripple detection · intracranial EEG · SEEG · closed-loop stimulation · threshold calibration · NREM sleep · F1 score · signal processing