One-Night Personalization Improves YASA-Target Wearable Sleep-Duration Estimation in Autistic Adolescents

Sleep disturbance is a common and actionable co-occurring medical concern for autistic young people, but population wearable algorithms may carry stable person-specific error. We tested whether one concurrent automated EEG-watch night can personalize total-sleep-time (TST) estimation on later nights. Public Simons Sleep Project summaries provided YASA-derived Dreem3 home-EEG TST as the target and EmbracePlus TST as the deployment predictor. A participant-weighted Huber model supplied population calibration. In five family-held-out folds, the first chronological calibration residual for each held-out participant was shrunk by an empirical-Bayes gate estimated exclusively from cross-fitted training-family residuals, then applied to strictly later nights. Among 88 autistic participants and 855 later nights, participant-balanced mean absolute error (MAE) fell from 60.73 to 53.40 minutes: personalized-minus-population difference -7.33 minutes (95% family-bootstrap CI -13.48 to -1.67). All 20 family-split seed averages were favorable. First-night residuals predicted later personal bias (Spearman rho = 0.610), while assigning the wrong participant's residual worsened MAE in 2,000 falsification repeats. Participant-mean concordance improved, but individual-night agreement limits remained about -147 to +145 minutes. On identical support, the YASA-target effect persisted (-7.77 minutes), whereas the Dreem-target effect was null (+1.08; 95% CI -2.46 to +5.10). One-night shrinkage therefore recovered a reusable, person-specific YASA-watch bias in this selected cohort but did not establish clinical accuracy or reference-independent generalization.