A residual deep learning framework for sleep apnea diagnosis from single lead electrocardiogram signals: An explainable artificial intelligence approach

Abstract

Sleep apnea is a common sleep disorder that occurs due to repetitive obstruction of the airflow in human body and affect human lives. Conventional machine and deep learning-based methods hardly have the transparency in the decision making and work as a “black-box”. To alleviate this problem, this article aims to present an explainable deep learning framework for diagnosis of sleep apnea using single channel electrocardiogram signals. The recorded electrocardiogram signals are preprocessed and converted into two-dimensional time-frequency images via continuous wavelet transform for necessary feature extraction followed by classification via residual neural network. The prime motive to employ time-frequency representation is to produce temporal and spectral information simultaneously. Also, time-frequency graphs are beneficial to analyze and represent non stationary signals possessing multiple time varying frequencies. The contributions of the proposed research are threefold: Firstly, the proposed framework not only diagnose sleep apnea in terms of classification accuracy, but also obtain the graphical explanation in a supervised manner. Secondly, relevance maps, generated through layer wise relevance propagation, are analyzed at each convolutional layer for multiclass sleep apnea diagnosis. Thirdly, an optimal explainable deep learning framework has been proposed to yield a fine-tuned and precise time frequency plot. Substantial experiments reveal that the proposed method achieves a significant accuracy of 98.16%, outperforming state-of-the-art sleep apnea diagnosis methods, along with specificity 99.05% and sensitivity 97.64%. Moreover, the method yields an accuracy of 97.52 % in another database. The presented topology can also be further used to other medical modalities in different biomedical systems.