Authors - Nitesh Pradhan, Aryan Baghla Abstract - Electrocardiogram (ECG) signals are effective indicators for detecting obstructive sleep apnea (OSA) due to their ability to reflect physiological changes associated with apnea events. EfficientApneaNet is a deep learning-based model for the detection of OSA from a single-lead ECG. In general, traditional approaches are reliable but exhaustive and costly; therefore, ECG-based methods are being sought. Earlier machine-learning methods, including Support Vector Machines and Random Forests, are prone to real data noise. Based on these, EfficientApneaNet joins previous advances in deep learning for further improvement in accuracy and robustness. The proposed architecture is powered by three novelties, namely: the ENBlocks, inspired by EfficientNet, Squeeze-and-Excitation blocks, and attention. ENBlocks make use of depthwise separable convolutions that reduce the computational complexity and amplify the efficiency of spatial feature extraction. The Squeeze and Excitation blocks carry out channel recalibration to focus on relevant patterns, while the attention mechanism underlines the critical temporal events within the ECG sequences. On the Apnea-ECG dataset, EfficientApneaNet realizes state-of-the-art performance with 91.47% accuracy, 85.92% sensitivity, and 94.91% specificity outperforming those of leading existing CNN-LSTM hybrids. It adopts Adamax optimization for stability while implementing the technique of cosine annealing for LR scheduling. The residual connections avoid gradient vanishing and explosion. Through ablation studies, it is confirmed that SE blocks and the attention mechanism are both essential to achieving high sensitivity and high specificity. In this respect, EfficientApneaNet would be considered a significant improvement in OSA detection, as it has successfully handled spatial and temporal complexities in ECG data.
