Obstructive Sleep Apnea (OSA) is the most common breathing-related sleep disorder, leading to increased risk of health problems. In this study, we investigated and evaluated the supervised machine learning methods to predict OSA. We used popular machine learning algorithms to develop the prediction models, using a dataset with non-invasive features containing 231 records. Based on the methodology, the CRISP-DM, the dataset was checked and the blanked data were replaced with average/most frequented items. Then, the popular machine learning algorithms were applied for modeling and the 10-fold cross-validation method was used for performance comparison purposes. The dataset has 231 records, of which 152 (65.8%) were diagnosed with OSA. The majority was male (143, 61.9%). The results showed that the best prediction model with an overall AUC reached the Naïve Bayes and Logistic Regression classifier with 0.768 and 0.761, respectively. The SVM with 93.42% sensitivity and the Naïve Bayes of 59.49% specificity can be suitable for screening high-risk people with OSA. The machine learning methods with easily available features had adequate power of discrimination, and physicians can screen high-risk OSA as a supplementary tool.
Keywords: Data Mining; Obstructive Sleep Apnea; Prediction; Supervised Machine Learning Methods.