A patient-specific classifier detected 11 of 12 held-out seizures but produced 75 pre-ictal alarms per hour in four SEEG patients

In a leave-one-seizure-out comparison of causal seizure-onset detectors, a patient-specific L2 logistic classifier trained on two seizures detected 11 of 12 held-out seizures in four SEEG patients (median latency 0 s; median AUC 0.994). Fixed line-length thresholds detected 3/12 and 0/12, respectively, with median AUC 0.876 for both. The classifier produced 75 pre-ictal false alarms per hour (22 alarms in 0.29 h), compared with none for the more conservative threshold. Five of its 11 detections occurred 5 to 10 s before clinical onset and may be early false alarms. No test reached p < 0.05; the exact paired sign-flip test for detection versus the conservative threshold gave p = 0.125, the minimum possible with four patients. In three held-out seizures, the interictal-calibrated score saturated and AUC was 0.5. The sample included only four patients and 12 seizures, with two training seizures per fold and about 88 s of pre-ictal data per seizure; onset times were clinical marks. All effects are descriptive, false-alarm rates are coarse, and the classifier may overfit.

SEEG · seizure onset detection · patient-specific classifier · line length · false alarms · leave-one-seizure-out · OpenNeuro ds004100