ApneaTraNet: A Deep Fusion Model for Detection of Sleep Apnea-Hypopnea Using ECG Signals

Obstructive sleep apnea-hypopnea (OSAH) poses significant health risks and traditional diagnosis methods are costly and hospital-based. This study presents an innovative deep learning framework for wearable solutions. Based on single-lead electrocardiogram (ECG) signals, the proposed model employs a ResNet34 architecture extended with bidirectional gated recurrent unit and transformer layers. This hybrid sequential fusion enables robust feature extraction, crucial for identifying distinct breathing events, such as apnea, hypopnea, and normal, within 5-second ECG segments. Our approach was evaluated on datasets from two publicly available benchmark datasets. Based on 10-fold cross-validation, the model achieved 93.79 % accuracy for classifying apnea, hypopnea, and normal sleep states. For the binary classification task of differentiating apnea from normal sleep, the proposed model achieved 97.19% accuracy. These performance measures indicate that the proposed model outperformed all existing studies for sleep apnea detection. This underscores the potential for automated OSAH detection using a wearable device, offering a more feasible and cost-effective alternative to traditional methods.