In 24 seizures from 8 patients, LOSO lowered false-alarm rates for two trained SEEG detectors, mostly due to more training seizures

An implanted closed-loop device can learn only from earlier seizures. Sci publication d77e8271 tested three SEEG detectors with leave-one-seizure-out (LOSO) training. We used its code without change and changed only the training seizures. Data: OpenNeuro ds004100, 8 patients, line-length features from 99d96b92, 24 test seizures that have at least one earlier seizure. We assume that run order is time order. Rules: CAUSAL (earlier seizures only), LOSO (all other seizures), and SIZEMATCH (the same number of seizures as CAUSAL, in any time order). B1 has no seizure training and gave the same result under all rules (15/24, 77.8 false alarms per hour). B2 (threshold from interictal and pre-ictal data) detected 12/24 at 19.4 FA/h under CAUSAL and 11/24 at 5.6 FA/h under LOSO. B3 (logistic regression) detected 22/24 under all rules, at 45.8 FA/h (CAUSAL) and 27.8 FA/h (LOSO). Mean held-out AUC for B3 was 0.899 (CAUSAL), 0.921 (SIZEMATCH) and 0.953 (LOSO). The direction was the same in all patients with a difference (3 of 8 for B2, 4 of 8 for B3), but the per-patient sign-flip tests were not significant (p = 0.25 and 0.125). Time order explains only 14 % (B2) and 31 % (B3) of the false-alarm gap. The remainder comes from the larger number of training seizures in LOSO. The gap is largest with one earlier seizure (B3: 87.5 vs 33.3 FA/h). With 3 or 4 earlier seizures, the rules gave near-same results. Thus LOSO false-alarm rates are optimistic for a device soon after implant. The LOSO arm reproduces all 72 per-seizure rows of d77e8271. Limits: 0.72 h of pre-ictal data, small false-alarm counts, 54 tests without correction.

seizure detection · SEEG · leave-one-seizure-out · causal training · false-alarm rate · closed-loop devices