DeepApneaNet: A Multistage CNN-Bi-LSTM Hybrid Model for Sleep Apnea Detection From Single-Lead ECG Signal

Abstract

Obstructive sleep apnea (OSA) is a critical sleep disorder that can lead to severe health complications and even death if left untreated. Early OSA detection through non-invasive methods, such as single-lead electrocardiogram (ECG) analysis, presents a promising approach for timely intervention. In contrast to the existing single-stage convolutional neural network (CNN) and bidirectional long short-term memory-based (BiLSTM) hybrid models, this letter presents DeepApneaNet, a novel end-to-end deep learning framework that combines multiple CNN-BiLSTM-based hybrid subnetworks in a cascaded manner to detect OSA from single-lead ECG signals. With the PhysioNet Apnea-ECG Database, our implemented framework is able to achieve the best per-segment accuracy, sensitivity, and specificity of 88.61%, 84.23%, and 91.04%, respectively. For per recording classification, our model achieved 94.29% accuracy, 100% sensitivity, and 83.33% specificity. Even though using minimal preprocessing and without any hand-crafted feature extraction, the performance of our model is still comparable to the state-of-the-art methodologies. The results indicate that segmenting hybrid models into smaller networks enhances the understanding of sequence dynamics. DeepApneaNet demonstrates significant potential as a practical solution for diagnosing OSA in real-world settings.