Answers are grounded in this paper and cite their sources.
Early-warning claims for electroencephalography (EEG)-based drowsiness detection rest on lead times, yet a lead time depends on how onset is operationally defined and is rarely tested against a null. We developed a benchmarking framework and applied it to seven onset estimators and 32 features from single-channel frontal EEG (17 participants, observer-verified drowsiness reports), with external replication in the DROZY dataset (13 sessions with an eligible vigilance-lapse anchor, of 20 retained). Estimator choice shifted onsets by minutes and changed detection coverage from 45% to 100% (Kendall’s W = 0.53; external W = 0.71; absolute-agreement intraclass correlation 0.05). In simulation, two-standard-deviation baseline rules flagged onsets in 77–88% of no-change trajectories; null-calibrated thresholds (about 4.6 standard deviations) restored a nominal 5% rate. No estimator showed a statistically detectable excess of onsets in the final 150 s before the anchor relative to spectrum-preserving surrogates, a consensus detector fired before pseudo-anchors as often as before the true anchor, and adding EEG to elapsed session time gave no detectable improvement (area-under-the-curve difference 0.01, 95% confidence interval −0.11 to 0.13). Higuchi fractal dimension and line length correlated weakly with measured electrooculography (|r| ≤ 0.15). We propose interpretability criteria, evidence levels and a reporting checklist for early-onset claims.
No takes yet. Share an insight, caveat, or question.
Abdi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: