ESSCNN-automatic sleep apnea detection throughpre-trained ResNet50 using electrocardiogram signal

Sleep apnea (SA) is caused due to sudden drops in oxygen levels which create stress in the body. SA is increasingly common in our society. SA is characterized by airway collapse when we are sleeping which is chronic in nature. Most of the time people are unaware that, they are undergoing this disorder and thus remain undiagnosed. Untreated SA may be the consequence of high blood pressure, arrhythmia, cardiovascular diseases, strokes, and heart attack which is mortal ultimately. To detect SA most of the research work has already been done using polysomnography (PSG). We propose a novel method to detect and classify SA as normal, borderline, and apnea using a single-channel ECG signal in a pre-trained ResNet50 convolution neural network. In our research work, we have reported pre-trained convolution neural network (CNN) ResNet50 for classification using Physionet apnea ECG for training and testing and further, we have also investigated the robustness of the model in the University College Dublin (UCD) sleep apnea database. The different solvers adam, rmsprop, and sgdm, and two different learning rates 0.001 and 0.0001 are used to get the optimum result in pre-trained ResNet50 to diagnose SA. The detection of SA using our proposed novel method has achieved outstanding performance and achieved 100% accuracy for both the training and testing phase. Resnet50 is a high-accuracy model. Comparison of performance of the model in state-of-art training loss, and validation loss implemented on both physionet apnea ECG data and UCD apnea data. We conclude that all the trials are showing 100% accuracy but there is a difference in the training loss and validation loss. For learning rate 0.001 and adam solver the model is showing the least training loss as well as validation. Consequently, for learning rate 0.0001 and sgdm solver the model is showing the highest training loss as well as Figure 1. Block diagram and process flow diagram of the proposed system Figure 1. Block diagram and process flow diagram of the proposed system validation. The proposed model is highly accurate and efficient in classifying all three classes normal, apnea, and borderline subject. In future work, we plan to do performance optimization and hyperparameter tuning using various pre-trained architectures for SA detection and its characterization for determination of severity The training and testing both could be done in the graphics processing unit to improve the training and testing time.