Four training-selected SEEG pairs matched all-pair AUC in eight patients but detected fewer held-out seizures at the fixed threshold

We tested whether four bipolar pairs selected from training seizures could retain held-out seizure-detection performance compared with all pairs and random subsets. We used archived line-length features from OpenNeuro ds004100 version 1.1.3 for eight patients and 32 seizures. In leave-one-seizure-out evaluation, pair selection, model fitting, and thresholds used only the interictal run and other seizures from the same patient. With logistic regression, top-4 pairs had mean AUC 0.931, compared with 0.927 for all pairs and median 0.833 for random-4 subsets. Top-4 AUC exceeded random-4 in all eight patients (exact paired sign-flip p = 0.008); top-4 and all-pair AUC did not differ clearly (+0.010, p = 0.48). At the threshold set to produce no false alarms on training negatives, top-4 detected 24/32 held-out seizures with 12.5 false alarms per hour of pre-ictal test data, while all pairs detected 29/32 with 24.0 per hour. The four-pair detector therefore had a different operating point on a similar ROC curve. The study includes eight patients, uses only 0.96 hours of pre-ictal clips to estimate false alarms, uses clinical onset marks, selects pairs within the same session, and does not model RNS hardware. Results do not establish that the selected pairs remain best months later or transfer to RNS electrodes.

seizure detection · stereo-EEG · SEEG · responsive neurostimulation · channel selection · line length · leave-one-seizure-out · AUC · false alarms · OpenNeuro ds004100