SEEG onset pairs increase in the last 4 s before clinical marks in 20 patients, but a per-seizure threshold rule does not outperform random pairs

Closed-loop SEEG devices and seizure detectors use clinical onset marks as truth. We tested if the SEEG change at the usual onset pairs starts before the mark. We used archived 1-s features of 77 seizures from 20 patients with 2 or more seizures (OpenNeuro ds004100, HUP; sci 3b7b4a7a): log 20-100 Hz power (HF) and log line length (LL) on bipolar pairs, z-scored against [-40,-25) s. For each seizure, we selected the onset pairs from the OTHER seizures of the same patient (top 10% of the mean z in [0,10) s). Thus a seizure does not select its own pairs. We wrote the plan before the analysis. Primary result: the change D = mean z in [-5,0) s minus mean z in [-20,-15) s was larger at onset pairs than at other pairs (HF median +0.43 z, 15/20 patients, exact Wilcoxon p = 0.004; LL +0.76 z, 18/20, p = 1e-5). Other pairs and random pairs of the same count did not change (Wilcoxon p >= 0.29). Post hoc per-second test with family-wise control (Wilcoxon max statistic over -25..-1 s): the onset-minus-other difference was significant from -4 s for both features; uncorrected tests were p < 0.05 from -8 s. A per-seizure threshold rule (3 s above the baseline maximum) found a change before 0 s in 33/77 seizures (HF), but random pairs also crossed before 0 s in 34% of seizures. Thus the rule did not separate onset pairs from random pairs (p = 0.11 HF, 0.24 LL). Interpretation: at the group level, clinical marks come after an electrographic change at the onset pairs. Detector scores that count detections in the last 4-8 s before the mark as false alarms can be biased. Limits: one center, human onset marks, only 15 s of baseline, post hoc per-second tests. The package has the code, the 79 input files with SHA-256 and S3 ETags, per-seizure tables, figures and reproduction commands.

SEEG · epilepsy · seizure onset · clinical onset marks · high-frequency power · line length · brain-computer interfaces · OpenNeuro