Detection of Sleep Apnea from Single-Lead ECG: Comparison of Deep Learning Algorithms

Abstract

Apnea is a prevalent sleep disorder which has detrimental impacts on human health and quality of life. Accurate automatic algorithms for the detection of sleep apnea are needed for analyzing long-term sleep data and monitoring and management of its side effects and consequences. Among different approaches for automatic detection of sleep apnea from biosignals, deep learning algorithms are of particular interest as, unlike conventional machine learning algorithms, they do not rely on expert crafted features. In this paper, we developed and evaluated a number of different deep learning models for the detection of sleep apnea from a single-lead electrocardiogram (ECG) signal. ECG R-peak amplitude and R-R intervals were extracted, and power spectral analysis was performed to align the R-peak amplitude and the R-R intervals in frequency domain. Convolutional neural network (CNN), long short-term memory (LSTM), bidirectional LSTM, gated recurrent unit, and deep hybrid models were implemented and analyzed. The performance of deep learning algorithms was evaluated on an apnea-ECG dataset of 70 recordings divided into a learning set of 35 records and a test of 35 records. The best accuracy, sensitivity, specificity, and F1-score on the test data were 80.67%, 75.04%, 84.13%, and 74.72%, respectively, with a hybrid CNN and LSTM network. The results show promise toward improved apnea detection using deep learning.