Meta-learning Based Obstructive Sleep Apnea Detection Using Single-Lead ECG Signals

As a respiratory syndrome correlated with some cardiovascular diseases, obstructive sleep apnea (OSA) not only destroys the quality of our sleep, but also induces a variety of major chronic diseases such as heart disease, and diabetes, and even causes sudden death during sleep. Many studies have been conducted on the classification of OSA from normal events by machine learning. However, we found that differences in patients caused by the individuality of the ECG patterns and variability in the ECG do not create optimal rules for OSA classification by ECG signals. It is necessary to reduce the impact of individual differences in classification. In this study, we propose a method based on meta-learning to detect OSA using a 2D time-frequency scalogram that is converted from a single-lead ECG signal. According to the experiment results, we achieved 68.10% accuracy, 69.19% sensitivity, 67.71% specificity, and a 0.694 F1 score. The results showed this method based on meta-learning using the Siamese network is feasible without being pre-trained by massive 2D scalogram representations.

I. Introduction

Obstructive sleep apnea (OSA) arises due to the total (apnea) or partial (hypopnea) repeated interruption of breathing during sleep and induces a variety of major chronic diseases. With the maturity of machine learning technology, more and more researchers are using machine learning to detect OSA from normal. We found that the classification of ECG signals is a challenging problem due to individual differences in patients caused by the individuality of the ECG patterns and variability in ECG waveforms by patients. This issue has no existence of optimal classification rules for ECG classification [1], [2]. Developing the most appropriate classifier that is capable of classifying OSA more accurately by reducing the impact of individual differences is a challenge in OSA classification.