In 19 SEEG patients, first-seizure maps selected top high-frequency onset pairs more often than interictal features

We tested if a 5-min interictal SEEG clip can select the bipolar pairs with the largest 20-100 Hz power increase at seizure onset (top 10% of pairs in each seizure) before a seizure is recorded. This is a target-selection step for responsive stimulation. We joined the per-seizure onset features of 3b7b4a7a with the interictal features of a1218d2f (19 patients, 72 seizures, 26-151 pairs each). We wrote the analysis plan before the results. A leave-one-patient-out logistic model trained on clinical SOZ labels (the a1218d2f settings) ranked the clinical SOZ well (AUC 0.83). But it ranked the onset pairs with median AUC 0.59 (12/19 patients above 0.5; p = 0.11). This was not different from the clinical SOZ itself (0.57; p = 0.52). The same 18 features, trained on onset labels of the other patients, gave AUC 0.67 (16/19; p = 0.0001). This was better than the SOZ-trained model (p = 0.012) and the clinical SOZ (p = 0.036). The onset map of the first seizure predicted seizures 2..n with AUC 0.71. It put the top onset pair in its 4 highest pairs in 0.33 of seizures. All interictal predictors did this in at most 0.13 of seizures (chance 0.06). Combining the interictal model with the seizure-1 map gave no gain. Limits: one center, 19 patients, a data-driven HF onset definition, and the assumption that run order is seizure order. The package contains the inputs with SHA-256 values, numpy-only code that reproduces the results byte for byte, tables and SVG figures.

stereo-EEG · epilepsy · seizure onset · interictal features · responsive stimulation · high-frequency activity