Detection of sleep apnea using deep neural networks and single-lead ECG signals

Abstract

Sleep apnea causes frequent cessation of breathing during sleep. Feature extraction approaches play a key role in the performance of apnea detection algorithms that use single-lead electrocardiogram signals. Handcrafted features have high computational complexity due to their large dimensions and are usually not robust. To cope with the mentioned problems, in the current paper, an automatic feature extraction method is developed by combining the Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) recurrent network. Also, the fully connected layers are utilized to distinguish apnea events from the normal segments. Then, the apnea-hypopnea index (AHI) is applied to discriminate apnea subjects from healthy ones. Finally, in order to assess the usefulness of the proposed method, some experiments are conducted on the publicly accessible Apnea-ECG and UCDDB datasets. The results based on the sensitivity (94.41%), specificity (98.94%), and accuracy (97.21%), indicate that our proposed method provides significant improvements compared to the other sleep apnea detection methods. Our model also achieves an accuracy of 93.70%, sensitivity of 90.69%, and specificity of 95.82% for UCDDB dataset. It can be inferred that using the deep-learning based algorithm for detecting apnea patients would help physicians in making a decision more accurately.