In 20 HUP Patients, Few-Pair Detectors Were Faster on SOZ Contacts, but Full Non-SOZ Sets Matched Them

A responsive-neurostimulation brain-computer interface (BCI) must detect seizures early with few contacts. We compared a causal line-length detector using bipolar SEEG pairs from clinical seizure onset zone (SOZ), NEAR, or FAR classes in OpenNeuro ds004100: 57 seizures from 21 patients, including 20 with SOZ labels. Latency was relative to the clinical onset mark, censored at +30 s; each seizure had its own baseline threshold. For random two-pair subsets, patient-median latency was 5.4 s for SOZ and 19.8 s for FAR (13/20 patients faster with SOZ; sign-flip p=0.011); detection by +10 s was 0.92 versus 0.22 (p=0.014). SOZ versus NEAR latency was 5.4 versus 10.0 s (11/16 faster; p=0.107). With all available pairs, SOZ and FAR latency was 4.3 versus 6.0 s (9/20 faster with SOZ; p=0.77), and SOZ versus NEAR was 3.1 versus 5.9 s (8/16 faster; p=0.40). Thus, in this sample, few-pair FAR detectors were slower and detected less often than SOZ detectors, while full non-SOZ sets performed about as early as full SOZ sets. Full-set detectors alarmed before the clinical onset mark in 11–21% of seizures; these alarms were not separated into early ictal changes and false alarms. Limits include clinical SOZ labels, short per-seizure baselines, no interictal false-alarm measurement, one feature, one center, and 57 of 104 planned seizures. The result does not establish prospective false-alarm performance or generalize beyond the studied setting.

SEEG · seizure detection · seizure onset zone · line length · responsive neurostimulation · intracranial EEG · detection latency · OpenNeuro