A Novel Multimodal Deep Learning Approach With Loss Function for Detection of Sleep Apnea Events

Abstract

Sleep apnea, a common and potentially serious sleep disorder, is characterized by repeated pauses in breathing during sleep. These pauses, known as apneas, can last from a few seconds to minutes and may occur multiple times per hour. This disorder poses significant health risks, including increased likelihoods of developing cardiovascular complications such as hypertension, heart disease, and stroke. Detecting sleep apnea accurately and efficiently presents several challenges, including variability in physiological signals among individuals and class imbalance for apnea events. In this study, we propose a novel deep-learning approach for automatic sleep apnea detection using physiological signals. Our method utilizes a convolutional neural network (CNN) with gated recurrent units (GRU) and an attention mechanism to capture both spatial and temporal information. We introduce a multi-domain feature extractor (MDFE) block with a Squeeze and Excitation (SE) block for enhanced feature representation and channel-wise attention. To promote model diversity and robustness, we further introduce a novel loss function that encourages the network's key modules to learn disparate feature representations. This approach is trained and validated on the Sleep Heart Health Study (SHHS) dataset, achieving high accuracy and robustness (AUC: 91.94, Accuracy: 83.87). We used the K-fold cross-validation technique to show the effectiveness of our method, suggesting its potential for real-time monitoring and personalized treatment strategies in sleep disorders.