Pre-analysis plan: participant-independent re-evaluation of physiological sleep-deprivation detection in the AdVitam Exp4 driving-simulator corpus

Pre-specified analysis plan, frozen on 22 October 2026 before any model was trained, for a re-analysis of the open AdVitam Exp4 corpus (63 drivers; mild sleep deprivation × rural/urban automated driving; ECG, EDA and respiration at 1000 Hz). The study tests how much of the published performance for detecting sleep deprivation from physiological signals depends on splitting data by segment rather than by participant, and reports participant-held-out, environment-held-out and joint hold-out results, label validity against sleep-tracker, KSS and PVT measures, and calibration.

The plan fixes the cohort, targets, quality-control thresholds, features, models, nested tuning, metrics, inference and directional hypotheses. Data seen before freezing are listed in the plan. Clarifications v1.1–v1.2 were made the same day, before any model was trained on real data, and are recorded in its change log (section 11). Later changes will be released as new versions of this record.

SHA-256 of ANALYSIS_PLAN.md: 3b5b25986e6145eb5e511fef46fffca3c015ebfc3c1505011b86efa849482c25