A30-14 Ensemble Learning-based Classification of Obstructive Sleep Apnea Severity Using Samsung Galaxy Watch4 Data: A Comparative Study With Polysomnography

Rationale Obstructive sleep apnea (OSA) is a common sleep-related breathing disorder characterized by recurrent upper airway obstruction during sleep, leading to intermittent oxygen desaturation and sleep fragmentation. Polysomnography (PSG) remains the gold standard for OSA diagnosis, but its high cost, need for trained personnel, and artificial laboratory environment limit accessibility and scalability. Wearable devices equipped with photoplethysmography (PPG) and accelerometer sensors provide an opportunity for noninvasive, continuous monitoring of physiological signals during natural sleep. This study aimed to classify OSA severity using biosignals from the Samsung Galaxy Watch4 and to validate ensemble learning models against PSG-based diagnoses. Methods We prospectively enrolled 150 adults with suspected sleep disorders who underwent overnight PSG at the Sleep Center of Wonju Severance Christian Hospital. Each participant wore Galaxy Watch4 devices on both wrists to enhance signal stability and reduce lateral bias. OSA severity was categorized into normal, mild, moderate, and severe groups according to the apnea-hypopnea index (AHI) derived from PSG. Features were extracted from accelerometer signals (AccX, AccY, AccZ), PPG signals (PPG, Compensated Infrared [Comp_IR], Compensated Red [Comp_RED]), heart rate (HR), HR variability (eSDHRState), and arousal-related events (eSDForceWakeUp). Four ensemble learning models—Random Forest, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost)—were trained using a 70:30 train-validation split. Model performance was evaluated by accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Results All ensemble models achieved robust classification performance, with mean accuracy of approximately 80%. XGBoost exhibited the best precision-recall balance, while LightGBM demonstrated superior generalization across OSA severity levels. AUC analyses showed clear discrimination between normal and severe OSA categories. Confusion matrix evaluation revealed minimal misclassification between adjacent severity groups, confirming the reliability of smartwatch-derived features in identifying OSA severity. Conclusions The ensemble learning approach using biosignals from the Samsung Galaxy Watch4 showed high concordance with PSG-based classification of OSA severity. This wearable-based method provides a scalable and noninvasive tool for early screening and longitudinal monitoring of sleep-disordered breathing. Future multicenter studies with larger populations are warranted to enhance model generalizability and establish clinical applicability. This abstract is funded by: None