Past-seizures-only channel selection gave the same SEEG detection sensitivity as leave-one-seizure-out in 8 ds004100 patients (24 test seizures, line-length detector, run order assumed as time order)

An implanted closed-loop device can learn only from earlier seizures. Offline studies often use leave-one-seizure-out (LOSO) cross-validation, which also uses later seizures. We tested whether LOSO makes a patient-specific SEEG detector look better than causal (past-only) training. We used archived line-length features from OpenNeuro ds004100 (8 patients, 24 test seizures with at least one earlier seizure). The detector and threshold were the same as in publication 99d96b92. Only the channel selection changed: the k bipolar pairs with the highest early ictal z, from earlier seizures only (CAUSAL) or from all other seizures (LOSO). Sensitivity was 0.71 vs 0.75 (k=1), 0.79 vs 0.79 (k=2), 0.71 vs 0.67 (k=4) and 0.63 vs 0.63 (k=8). Over 96 paired decisions, the two rules gave different results for 2 seizures only, one in each direction. Per-patient margin differences had no consistent sign (all sign-flip p >= 0.25). One earlier seizure was enough to select 1 or 2 good pairs. All rules had high pre-ictal false-alarm rates (67 to 138 per hour), because the threshold uses one 5 min interictal clip. Limits: small sample, short clips, run order assumed as time order, one feature and one selection rule. Code, features, results tables and a figure are in the package.

SEEG · seizure detection · leave-one-seizure-out · causal training · channel selection · line length · epilepsy