Identifying episodes of sleep apnea in ECG by machine learning methods

The paper is devoted to the application of machine learning methods for computerized sleep apnea detection based on single lead electrocardiographic signal (ECG). In order to explore the possibilities of machine learning for ECG-based apnea detection, the Apnea-ECG database provided by the PhysioNet resource was used in this study. A distinctive peculiarity of this database is that it contains the annotations for each minute of each recording indicating the presence or absence of apnea at the current moment of time. To evaluate the effectiveness of ECG derived parameters for sleep apnea detection, 80 ECG segments of 10 minutes duration, annotated as apnea, and 73 ECG segments of the same duration, annotated as normal sleeping were extracted and investigated. The purpose of this work is to define and compare the informative features for identifying episodes of sleep apnea in ECG by heart rate variability analysis, as well as to choose the classification method that provides the highest accuracy for this task. The time-domain, frequency domain, spectral-temporal and wavelet features are considered. Using these feature sets, the performances of a number of classifiers based on decision trees, discriminant analysis, logistic regression, support vector machines, variations of k-nearest neighbors' method, and ensemble learning, were determined. Based on this, a combination of features and classifiers are proposed that provides the highest accuracy of sleep apnea episodes recognition in single lead ECG. The choice of model options for the best performing classifiers was investigated.