Comparing Machine Learning Models for Sleep Apnea Detection using Physiological Data

Sleep apnea is a common sleep disorder that remains underdiagnosed due to the cumbersome nature of inconveniences related to accessibility and cost of traditional modalities for diagnosis, such as polysomnography (PSG). Machine learning (ML) is a burgeoning area for scalable, data-driven interventions for the early detection of sleep apnea, and the most viable physiological data collection methods are the wearables and smart devices rapidly gaining popularity. This paper compared supervised ML classifiers - Logistic Regression, Support Vector Machine (SVM), Random Forest, and XGBoost using a CSV dataset with Features; including age, BMI, oxygen saturation, and heart rate. Unlike most datasets, this dataset was a combination of real and synthetic samples, and involved some preprocessing efforts using median imputation, SMOTE, and normalisation due to imbalance in dataset. Exploratory Data Analysis (EDA) was completed for correlation identification. Each model was assessed and compared using established metrics: accuracy, precision, recall, F1-score, and ROC-AUC. The results yielded overall evidence that predictive performance would be maximised through use of ensemble methods and XGBoost broadly maintained favourable predictive performance with accuracy of 98.25%. It was concluded that ML is a timely complement to build and sustain robust, scalable options for data-driven diagnostic systems to identify sleep apnea.