In 15 held-out SEEG patients, a 5-second onset map predicted later seizure onset pairs, but prior-seizure averaging did not improve prediction

Question: does the first-seizure SEEG target map of sci 3b7b4a7a replicate in patients that the original analysis did not use, and how much of the seizure must the system record? Data: per-run 1-s features (sci 77969975) of OpenNeuro ds004100 v1.1.3 SEEG ictal runs. We used 15 held-out subjects (43 seizures) and the original 20 subjects (82 seizures). The plan was written before the analysis. Target: top 10% of bipolar pairs by mean 20-100 Hz power z-score in [0,10) s after the clinical onset mark of each later seizure. Results (held-out, Wilcoxon across subjects): the seizure-1 onset map gave median AUC 0.745 (14/15 subjects above 0.5, p=0.0001). It was better than clinical SOZ labels (AUC 0.53; difference +0.16, 14/15, p=0.0001). The onset-map correlation was higher than the sham-map correlation (+0.44, p=0.0006). The mean map of all prior seizures was not better than the seizure-1 map (-0.005, 6/13, p=1.0); thus this original result did not replicate. A 5-s seizure-1 map gave AUC 0.731 against 0.745 for 10 s. 1-3 s maps gave 0.62-0.66, above chance but worse than 10 s (p<0.01). Line length gave similar results. Limits: one center, small n, run order assumed to be seizure order, target from the same feature (repeatability, not outcome), windows aligned to the clinical onset mark (not causal detection), and a sham map slightly above chance (0.55). Code, features, hashes, results and SVG figures are in the package.

SEEG · epilepsy · seizure onset · onset zone · intracranial EEG · held-out replication · AUC · responsive stimulation