Allocating a Simulated Review Budget to Sleep–Wake Disagreements: Exploratory Analyses of Two Public Datasets

Automated sleep staging supports summary sleep measures, but the best use of a limited review budget remains uncertain. We present two development analyses of simulated review against expert sleep-stage references. In BOAS, training-derived quantities from 40 participants were compared across review policies in 20 previously inspected participants, using existing PSG and wearable headband predictions. At a 10% budget, mean absolute error in sleep time restricted to reference-scorable epochs was 7.725 minutes without review, 6.951 with random review, 4.475 with sleep-wake disagreement, and 6.250 with a conditional-risk rule. A reference-mask-conditioned bias correction yielded 11.888 minutes and is not a deployable automatic baseline. A separate Sleep-EDF feasibility probe used the same YASA model on two EEG leads from six selected participants; five were scored after one HTTP failure. Corresponding mean errors were 107.6 minutes without review, 95.7 with random review, 87.3 with model entropy, and 77.2 with lead disagreement. The BOAS source model's participant-level out-of-fold status and Sleep-EDF pretraining overlap are unverified; the latter also raised an unresolved scikit-learn version warning. The analyses use different inputs, selection rules and recording durations, and neither is independent validation. Perfect reference substitution was simulated, conditional outcomes are not full-night clinical total sleep time, and no human review time was measured. These exploratory findings do not establish clinical utility or generalizable superiority.