Early Detection of Sleep Disorders Using Ensemble Boosting and Classical ML with Lifestyle Data

Abstract

Sleep disorders are increasingly prevalent in modern society, significantly impacting quality of life, productivity, and physical and mental health. This study distinguishes itself by evaluating six machine learning algorithms—Support Vector Machine (SVM), Naïve Bayes, K-Nearest Neighbor (KNN), XGBoost, CatBoost, and LightGBM—using a comprehensive preprocessing pipeline that ensures balanced, normalized, and leakage-free training data, enabling a robust comparison of classical and ensemble boosting models for sleep disorder prediction. Data preprocessing included handling missing values, encoding categorical features, normalization, class imbalance correction via Random Oversampling, and stratified train-test splitting. Models were optimized through hyperparameter tuning with GridSearchCV and evaluated using accuracy, precision, recall, F1-score, and multiclass ROC-AUC metrics. Feature importance analysis revealed that Age, Occupation, Heart Rate, Sleep Duration, and Daily Steps were the most influential predictors, highlighting the interpretability of the models. LightGBM achieved the highest predictive performance with 70.0% accuracy and a ROC-AUC of 0.736, followed by Naïve Bayes (67.5% accuracy, ROC-AUC 0.784) and XGBoost (65.0% accuracy, ROC-AUC 0.728). McNemar’s test indicated no statistically significant difference in performance among models, confirming prediction stability. Models also differed in computational efficiency, with Naïve Bayes being the fastest and CatBoost the slowest, reflecting algorithmic complexity. These findings suggest that ensemble boosting algorithms, particularly LightGBM and XGBoost, alongside classical models like Naïve Bayes, provide effective, interpretable, and reliable tools for early detection of sleep disorder risk, with potential applications in wearable devices, digital health platforms, and clinical monitoring systems.