WIVIDOSA-Net: Wigner–Ville distribution based obstructive sleep apnea detection using single lead ECG signal

Abstract

Obstructive sleep apnea (OSA) is a serious condition causing intermittent breathing stops during sleep. Currently, it is diagnosed with polysomnography (PSG), which is costly and sometimes uncomfortable. Researchers are now exploring the use of electrocardiogram (ECG) signals as a potential alternative for diagnosing OSA. Here, we have proposed a novel deep learning model (DLM) to detect OSA using smoothed Wigner–Ville spectrograms (SWVSs) of ECG signals. The PhysioNet Apnea ECG Database (70 full-night ECG recordings) is used to validate the model performance. The proposed model first converted the per-minute ECG signals into WVSs and smoothened them using Savitzky–Golay (S–G) filtering. Then, SWVSs were fed as input to our newly developed DLM named WIgner–VIlle Distribution-based Obstructive Sleep Apnea convolutional neural network (WIVIDOSA-Net) as well as other standard pretrained ResNet-18 and ResNet-50 for comparison. The WIVIDOSA-Net model achieves an average classification accuracy of 90.09%, specificity of 91.12%, and sensitivity of 87.40% when evaluated using a tenfold cross-validation method. The proposed model extracts high-resolution spatial and temporal information, making the pipeline very effective in discriminating OSA episodes from normal. Therefore, it exhibits superior performance in comparison to all current state-of-the-art approaches, with a reduced computation burden due to its limited number of learnable parameters.