A transparent AI assurance and benchmarking framework for EEG seizure detection on TUSZ seeded with a reproducible gradient-boosting ensemble.

Reliable evaluation is essential for clinically usable automated electroencephalography (EEG) seizure detection. Although the TUH EEG Seizure Corpus (TUSZ) is the de facto benchmark, studies often use non-canonical splits, mix patients across partitions, or evaluate on seizure-only segments, making results hard to compare. We release an open-source, transparent benchmark for offline binary seizure detection on TUSZ v2.0.3 that strictly follows the official Train/Dev/Eval partition and evaluates in continuous EEG. Each 60-s window (15-s stride) is mapped to expert features and classified by three CatBoost ensembles: a within-window base model and two temporal-context models (full and high-sensitivity). We cast operating-point selection as an alarm-budget decision problem: on the Development set, we jointly choose the score threshold and event post-processing to maximize event sensitivity while enforcing specificity ≥ 0.93 and false alarms ≤ 0.69 per hour, then freeze this decision policy and apply it once to the held-out Eval set. On Eval, we obtain AUROC 0.92 and balanced accuracy 0.83 at the window level and event sensitivity 0.75 at 0.68 false alarms/hour (time-based positive predictive value [PPV] 0.57; recall 0.76). We provide the protocol and results as a reproducible baseline and clinically oriented benchmark for comparing past and future models.