In HUP SEEG patients, a second seizure improves held-out onset-map AUC, but later seizures add little and do not improve top-pair capture

In 17 patients from one center in OpenNeuro ds004100, we assessed whether averaging more recorded seizures improves a 20–100 Hz SEEG onset map for a held-out seizure. We used archived per-second power and line-length features, clinical onset marks, and run order as seizure order. For each held-out seizure, we averaged maps from all subsets of k other seizures (k = 1 to n−1); outcomes were AUC for the held-out seizure’s top 10% of pairs and whether its top pair appeared among the map’s top four pairs (hit4). For power, median AUC gain was +0.041 from k=1 to 2 (17 patients, 15 up, p=0.0001), +0.010 from 2 to 3 (10 patients, 9 up, p=0.006), and +0.004 from 3 to 4 (10 patients, 7 up, p=0.065). In the 10 patients with five seizures, median AUCs were 0.73, 0.82, 0.85, and 0.86. Mean hit4 was 0.31, 0.34, 0.39, and 0.38; the k=1 to 2 and k=1 to 4 changes were not significant (p=0.20 and p=0.30). Line length showed AUC gains in the same direction, while hit4 did not increase. In a chronological check, mean AUC rose from 0.61 to 0.72 for k=1 to 2 (p=0.008). Clinical seizure-onset-zone labels gave median AUC 0.56 for the same targets. These results indicate that most AUC benefit came from the second recorded seizure; additional seizures improved the full pair ranking but did not make a four-pair target more likely to contain the held-out top pair. Seizure-to-seizure change in the top pair is a probable cause. This single-center, non-preregistered analysis used clinical onset marks, treated run order as seizure order, defined targets from the same feature rather than a surgical outcome, and included only 10 patients with k≥3.

stereo-EEG · seizure onset · epilepsy · onset mapping · responsive neurostimulation · held-out prediction · calibration seizures · line length · high-frequency power