Obstructive Sleep-Apnea Detection using Signal Preprocessing and 1-D Channel Attention Network

Abstract

The alarming increase in the number of people affected by Obstructive sleep apnea is a point of concern for the medical experts as well as the whole populace. The discontinuous breathing due to upper airway blocking may lead to various cardiovascular and neurological complications such as hypertension and stroke. Polysomnography and manual scanning of ECG signals are time-consuming and expensive techniques. Thus, a need arises for exploring the alternatives to these techniques. The existing literature highlights the efficacy of machine learning and deep learning techniques in the detection of sleep apnea using ECG signals. The comparison of 2-Dimensional and 1-Dimensional convolution neural network-based models stipulate that 2-Dimensional models are better in efficacy, whereas of 1-Dimensional models are more adaptive and compact. Therefore, in this study, the authors looked for a trade-off and designed a novel 1-Dimensional architecture ’1-D Channel Attention Convolution Neural Network’ for the detection of sleep apnea. This compact architecture achieves better accuracy than other 1-Dimensional models with a limited number of parameters. First, it uses the Savitzky-Golay filter for noise suppression, thereafter it uses the smoothened signals for classification. The model utilizes channel attention layers to refine the intermediate feature descriptors before passing them deeper into the network. Thus, simultaneously reducing the need of implementing a deep neural network. The 1-D Channel Attention Convolution Neural Network reports the highest average accuracy of 93.01%, specificity of 93.10%, and sensitivity of 92.93% for sleep apnea detection. The results obtained prove that the proposed model outperforms other state-of-the-art 1-D convolution neural network architectures.