In 20 SEEG patients, learned channels did not detect seizures earlier than clinical SOZ channels

We compared seizure detection from 1, 2, 4, or 8 bipolar pairs selected from a patient’s other seizures (LEARNED), clinical seizure-onset-zone (SOZ) pairs, random pairs, all implant pairs, or the best pairs of the test seizure (ORACLE). We used archived 1-s high-frequency power (hp) and line-length (ll) features from 77 seizures in 20 patients at one center (HUP; OpenNeuro ds004100 v1.1.3). With hp and k = 4, LEARNED latency was 6.19 s and SOZ latency was 5.94 s (difference +0.26 s; 7/20 patients faster with LEARNED; p = 0.84). Both were about 3 s faster than random pairs. LEARNED and SOZ were not different at any tested k or with ll. LEARNED was 3.2–6.1 s slower than ORACLE across k = 8–1, consistent with the best pairs changing between seizures. More channels improved detection speed but increased sham detections; all pairs detected at −0.4 s and had a sham detection in 46% of seizures. A causal selection using only earlier seizures (57 test seizures) was also not different from SOZ (−1.02 s; p = 0.50). Limits include one center, clinical onset marks and SOZ labels, short pre-ictal sham data, 1-s resolution, one detector threshold, and run order used as seizure order. The analysis plan was written before results were computed. See the main text for methods, full results, and limitations.

SEEG · seizure detection · responsive neurostimulation · channel selection · intracranial EEG · epilepsy