In 20 patients, SEEG seizure maps were more repeatable at 16–20 s than 0–4 s after the clinical onset mark

Question: a responsive-stimulation BCI that learns its targets from recorded seizures must choose a time window after seizure onset. Which 2-s window after the clinical onset mark gives the SEEG seizure map that is most repeatable across the seizures of one patient? Data: archived bipolar-pair features of sci 3b7b4a7a (OpenNeuro ds004100 v1.1.3, HUP), 20 patients, 77 seizures. Each pair is z-scored against [-40,-25) s. Repeatability is the median Spearman rho between the maps of two seizures of one patient. The plan was written before the maps were computed. Primary result (log 20-100 Hz power): median rho was 0.17 in [0,4) s and 0.53 in [16,20) s. E was larger than L in 3 of 20 patients (exact Wilcoxon p=0.002). Line length gave the same result (0.19 vs 0.47, p=0.00013). The [0,2) s map was not different from the pre-onset reference (rho 0.04, p=0.26). Repeatability started at 2-4 s and was highest at 10-20 s. A seizure-1 map from 4-20 s predicted the top-10% pairs of later seizures with median AUC 0.61-0.80; the [0,2) s map gave 0.53-0.60. Exploratory: late maps agreed with the clinical SOZ labels as well as early maps did (SOZ AUC about 0.7; [0,4) vs [16,20) s p=0.62). Limits: clinical marks can be late; features end at +20 s; late repeatability can partly come from a stereotyped spread; one center; 20 of 35 eligible patients. The package contains code (numpy and Pillow), inputs with checksums, results, figures and reproduction steps.

stereo-EEG · seizures · repeatability · responsive stimulation · seizure onset