Brief digital sleep questionnaire powered by machine learning prediction models identifies common sleep disorders

Introduction We developed and validated an abbreviated Digital Sleep Questionnaire (DSQ) to identify common societal sleep disturbances including insomnia, delayed sleep phase syndrome (DSPS), insufficient sleep syndrome (ISS), and risk for obstructive sleep apnea (OSA). Methods The DSQ was administered to 3799 community volunteers, of which 2113 were eligible and consented to the study. Of those, 247 were interviewed by expert sleep physicians, who diagnosed ≤2 sleep disorders. Machine Learning (ML) trained and validated separate models for each diagnosis. Regularized linear models generated 15–200 features to optimize diagnostic prediction. Models were trained with five-fold cross-validation (repeated five times), followed by robust validation testing. ElasticNet models were used to classify true positives and negatives; bootstrapping optimized probability thresholds to generate sensitivities, specificities, accuracies, and area under the receiver operating curve (AUC). Results Compared to reference subgroups, physician-diagnosed sleep disorders were marked by DSQ evidence of sleeplessness (insomnia, DSPS, OSA), sleep debt (DSPS, ISS), airway obstruction during sleep (OSA), blunted circadian variability in alertness (DSPS), sleepiness (DSPS and ISS), increased alertness (insomnia) and global impairment in sleep-related quality of life (all sleep disorders). ElasticNet models validated each diagnosis with high sensitivity (80–83%), acceptable specificity (63–69%), high AUC (0.80–0.85) and good accuracy (agreement with physician diagnoses, 68–73%). Discussion A brief DSQ readily engaged and efficiently screened a large population for common sleep disorders. Powered by ML, the DSQ can accurately classify sleep disturbances, demonstrating the potential for improving the sleep, health, productivity and safety of populations.