Analysis of Machine Learning Algorithm for Sleep Apnea Detection Based on Heart Rate Variability

Abstract

Sleep apnea is a common problem with health implications ranging from excessive daytime sleepiness to serious cardiovascular disorders. The method for detecting and measuring sleep apnea is through breathing monitoring (polysomnography), which is time consuming and relatively expensive. Cardiovascular which is closely related to heart performance activities allows the use of electrocardiogram (heart rate variability) features to detect sleep apnea. This study aims to compare the results of sleep apnea detection using several machine learning algorithms. A total of 2,445 data were divided into 1,834 data as learning sets and 611 data as test sets. Evaluation of 10-fold cross-validation using all HRV features shows that neural network algorithm has the best performance compared to decision tree algorithm, k-nearest neighbor, and support vector machine with an accuracy rate (82.44% in the learning set, 79.21% in the test set consecutively), precision (85.54% and 82.70%), f-measure (87.70% and 85.67%), and AUC (0.867 and 0.832). Based on the results of performance testing using only selected HRV features (CVRR, HF, SD1/SD2 Ratio, and S-Region), the K-Nearest Neighbors, Support Vector Machine, and Neural Network algorithms experienced a decrease in performance. The use of all HRV features is recommended compared to only using selected HRV features, so it can help detect the presence/absence of sleep apnea much better.