A Contactless Approach for Accurate Sleep Stage Detection Using the Sleepal AI Lamp

To validate the accuracy and clinical feasibility of a noncontact radar system (Sleepal AI Lamp) for assessing sleep architecture and staging. Although noncontact radar sensors can capture respiratory and motor mechanics without physical attachment, robust validation in large-scale, pathologically diverse cohorts remains scarce. Approach: Nocturnal respiratory and body movement signals were assessed in 1,022 adults using simultaneous measurements from a frequency-modulated continuous-wave (FMCW) bedside radar and gold-standard PSG. The cohort included individuals with normal sleep profiles and patients with varying severities of obstructive sleep apnea (OSA). A frequency-augmented deep learning architecture interpreted these respiratory and movement signals. Main results: In the independent validation set, binary sleep-wake classification achieved 92.7% accuracy, a Cohen's kappa of 0.792, 96.4% sleep sensitivity, 80.8% sleep specificity, and a macro-averaged F1 score of 0.896. Four-stage classification (Wake, Light, Deep, and rapid eye movement (REM) sleep) achieved 77.2% accuracy, a Cohen's kappa of 0.677, a macro-averaged F1 score of 0.768, and balanced accuracy/macro-recall of 77.4%, with stage-specific F1 scores of 0.839, 0.740, 0.691, and 0.802 for Wake, Light, Deep, and REM, respectively. Significance: The findings support the validity of the 60 GHz radar-based system for automated assessment of sleep macrostructure. By capturing stage-specific respiratory and movement phenotypes without physical attachment, this technology may provide an unobtrusive tool for home-based screening, longitudinal recovery tracking, and population-level sleep epidemiology.