Treatment coverage of concentrated 80–150 Hz SEEG power gave a small and uncertain Brier gain in 34 patients

We tested whether treatment coverage of concentrated 80–150 Hz power adds surgical outcome information beyond clinical coverage measures. We used 34 patients from OpenNeuro ds004100 version 1.1.3, with two 60-second SEEG clips per patient. Pair weights measure the power fraction in the six largest one-second windows, minus 0.1. They do not count detected epileptic events. The baseline used treated-pair fraction and coverage of seizure-onset pairs. We added excess-concentration coverage above treated-pair fraction to a fixed penalized logistic model. Each test patient was excluded from training, including all bootstrap copies. Mean leave-one-patient-out Brier error fell from 0.266740 to 0.262767, a gain of 0.003972. A 1,000-draw stratified patient bootstrap gave a 95% percentile interval of −0.008259 to 0.021858. Single-clip and strict-outcome checks gave mixed gains. A separate post hoc leave-pair-out audit gave AUC 0.541667 for the baseline and 0.576389 for the added model. The primary result does not establish useful prediction. Fresh source extraction reproduced all 68 power files exactly. Independent calculations and offline replay passed. The package preserves code, features, outcomes, bootstrap counts, plans, versions, hashes, retrieval recipes, results, figures, and limits.

stereoelectroencephalography · epilepsy · surgical outcome · high-frequency power · time concentration · patient-level prediction · Brier score