Advancing Sleep Apnea Detection with CINet: Convolution and Interactive Learning Neural Network for Analyzed ECG Signal Data

Sleep apnea, characterized by irregular breathing during sleep, can lead to severe health complications such as cardiac arrhythmia, myocardial infarction, and heart attack if left undetected. Early detection is crucial, yet the current standard method, polysomnography (PSG), is costly and inefficient. Advances in technology allow for the automatic classification of sleep apnea using electrocardiogram (ECG) signals. By processing these signals into 2D formats, image classification neural networks can be effectively utilized to improve detection accuracy and efficiency. In this study, ECG signals were frequency-filtered and segmented into noise-free epochs. Each 60-second epoch was then transformed into a spectrogram. We propose a spectrogram-based convolutional and interactive learning neural network (CINet) to recognize sleep apnea from ECG signals. The model employs a weighted odd-even convolutional branching structure to facilitate interaction between parallel blocks, thereby improving local-global feature extraction. The proposed model achieved an accuracy of 96.32% in classifying apnea and normal ECG segments. To demonstrate the performance gain of the proposed model, we compared its classification performance with 1D CNN, Inception-ResNet-V2, VGG-19, ResNet152V2, and similar classifiers in the literature.