Machine Learning-Based Prediction of Sleep Disorders from Lifestyle and Physiological Data: A Cross-Occupational Study

Abstract

Sleep disorders are increasingly recognized as critical public health concerns, particularly among working populations where occupational stress, lifestyle factors, and physiological imbalances intersect. This study explores the predictive capacity of machine learning models, including Random Forest, Support Vector Machine (SVM), and XGBoost to identify sleep disorders (None, Insomnia, and Sleep Apnea) using a dataset comprising demographic, occupational, lifestyle, and physiological variables. The dataset, drawn from 400 individuals, was preprocessed through normalization, one-hot encoding, and SMOTE to address class imbalance. Feature selection was conducted using correlation analysis, RFE, and Random Forest importance scores. Models were trained with stratified sampling and optimized using 5-fold cross-validation. XGBoost outperformed the others with an accuracy of 0.90 and an F1-score of 0.88, followed by Random Forest (0.875, 0.86), while SVM lagged (0.825, 0.71). Confusion matrix analysis revealed consistent misclassification between Insomnia and Sleep Apnea, reflecting overlapping symptomatology and low feature correlation. Occupational analysis showed that manual laborers exhibited higher stress levels and shorter sleep durations, particularly those with insomnia. These findings highlight the value of integrating occupational and physiological data into predictive modeling and underscore the potential of ensemble learning methods in health informatics. This study supports the development of early detection systems for sleep disorders tailored to occupational risk profiles.