Supervised classification of sleep events suitable for a minimized home-operated recording system

In this paper, we present a novel approach that uses a non-invasive recording device based on the respiratory rate, SpO2 (oxygen saturation), and heart rate (HR) to represent respiratory effort and function. We also use these signals to classify apnea, arousal, and hypopnea events. In addition to event classification, proposed machine learning models demonstrate the potential to estimate the heterogeneous Apnea-Hypopnea Index (AHI), which plays a pivotal role in designing the proposed novel recording device based on the respiratory rate, SpO2 (oxygen saturation), and heart rate (HR) signals. We based our recording device on the APPLES dataset, a publicly available dataset from which a set of 120 patients with sleep apnea who underwent overnight polysomnography recordings were used. We extracted time and frequency domain features from the respiratory signals after using wavelet decomposition and denoising of the abdomen effort signal. We trained machine learning models, such as support vector machine (SVM) and gradient boosting algorithm (XGBoost), using grid search for hyperparameter optimization. The Receiver Operating Characteristic (ROC) curves showed an overall area under the curve (AUC) of 0.89 for the classification task. Our approach achieved an accuracy of 0.8673, and an F1-score of 0.8670 using the XGBoost model. Additionally, a separate test set resulted in an accuracy of 0.8343 and an F1-score of 0.8341 for the XGBoost model. We compared our approach with existing methods using different vital signals, such as SpO2 and HR, and showed that our approach achieved superior performance in terms of AUC values for apnea (0.9), arousal (0.88), and hypopnea (0.82) using SpO2 signals; and apnea (0.92), arousal (0.91), and hypopnea (0.86) using HR signals.