Hybrid Residual Neural Network for Obstructive Sleep Apnea Detection Using ECG Scalogram.

The electrocardiogram (ECG) has emerged as a viable alternative to polysomnography (PSG) for the detection of obstructive sleep apnea (OSA). Given the limited feature information of ECG signals in a single domain, this study proposes an adaptive threshold denoising synchrosqueezed wavelet transform (ADSWT) algorithm to extract high-resolution time-frequency domain features of ECG signals. The ADSWT algorithm enhances the time-frequency resolution, enabling better extraction of key features while minimizing noise interference, and generates an ECG scalogram as input for the network. Additionally, we propose a hybrid residual neural network (HRN-Net) for the automatic classification of OSA. The HRN-Net is designed with a dynamic feature extraction module to parse complex local features and enhance the network's generalization capability, as well as a dependency modeling module to capture the relationships between global contextual information for better understanding and interpretation of the ECG signals. All-night PSG recordings from both public and private datasets are used to validate the proposed framework. The results show that the framework achieves an accuracy of 0.942, a sensitivity of 0.926, a specificity of 0.959, and an F1 score of 0.942 on the public dataset, and an accuracy of 0.946, a sensitivity of 0.921, a specificity of 0.972, and an F1 score of 0.945 on the private dataset. These results indicate that the proposed framework offers high accuracy in automatic OSA classification and has significant potential to aid clinical decision-making.