Scalp EEG seizure warning in children under a pre-specified, locked evaluation: confound audit, structured chance levels and a negative result

Reported performance of scalp electroencephalographic (EEG) seizure prediction is often inflated by lenient evaluation and by the vigilance-state, clock-time and admission-trend structure of monitoring data. We asked how much short-horizon warning survives a pre-specified, locked evaluation. Continuous video-EEG from 21 monitored children (1315 h) formed the development cohort. A statistical-process-control warning method, evaluated under a written protocol, detected 44% of evaluable seizures at 0.48 false alarms per hour, no better than a random predictor (51%, p = 0.86); its statistic moved three times more with vigilance state than with seizure approach. We re-designed it as a state-conditioned, directional scheme on seven global EEG features, froze it with a hash-recorded configuration and time-stamped pre-specification, and estimated its performance by nested leave-one-patient-out cross-validation: 15 of 45 seizures (33%) at 0.47 false alarms per hour, above analytic (21%, p = 0.034) and permutation chance, but a predictor using only clock time, vigilance state and days since admission reproduced most of this, and the signal was confined to the last 30 min. Applied once to two validation cohorts (a second batch from the same unit and the public CHB-MIT database; 101 evaluable seizures), the frozen scheme detected 19 seizures (19%) at 0.60 false alarms per hour against 26% chance (p = 0.96); no secondary endpoint was met, and with the child as the unit of analysis no cohort was above chance. Scalp-EEG warning should be evaluated under pre-specified rules against structured chance levels; the frozen scheme and framework are offered as a reference.