Predicting Polysomnography‑derived Sleep Efficiency and Total Sleep Time Using Clinical Questionnaires: A Machine...

We developed and evaluated machine learning models to predict polysomnography (PSG)-derived sleep efficiency and total sleep time using clinical questionnaire-based assessments, with the aim of developing accessible non-invasive screening tools that can estimate objective sleep parameters in the absence of PSG.

Predictive models were developed using demographic data and nine validated questionnaires from 1,135 participants. Three gradient boosting algorithms-Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LGBM), and eXtreme Gradient Boosting (XGBoost)-were compared for binary classification of sleep efficiency (≥ 85% vs. < 85%) and four-class prediction of total sleep time (< 5, 5-7, 7-9, and ≥ 9 hours). Model performance was evaluated using area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC) metrics.

CatBoost consistently outperformed LGBM and XGBoost. For sleep efficiency classification classification, CatBoost achieved the highest performance (AUC = 0.7342; PRC = 0.6957). In total sleep time prediction, CatBoost demonstrated superior accuracy (weighted AUC = 0.6772; weighted PRC = 0.5384; micro-AUC = 0.8195). Inclusion of formal sleep disorder diagnoses provided negligible improvements.

Validated self-report instruments effectively captured clinically meaningful determinants of sleep efficiency and total sleep time. Machine learning models, particularly CatBoost, demonstrated robust performance using questionnaire-derived features alone, highlighting the central roles of age, insomnia severity, and daytime sleepiness. These findings support the feasibility and clinical utility of accessible, questionnaire-driven machine learning approaches for large-scale screening, early risk stratification, and decision support in sleep medicine.