ExFormer: Exploring Convolution and Transformer Based Model for Sleep Apnea Recognition

Accurate detection of obstructive sleep apnea (OSA) through electrocardiogram (ECG) signals is essential for timely diagnosis and intervention, as untreated OSA can lead to serious health complications such as cardiovascular disease and cognitive impairments. The ability to analyze ECG signals for early identification of OSA provides a non-invasive, efficient means of facilitating prompt treatment, improving patient outcomes, and managing the condition. In this context, the present study introduces ExFormer, a novel CNN-Transformer model specifically designed for the detection of OSA using single-lead ECG data. The model employs a multiperspective convolutional network (MCN) for efficient extraction of salient features from short sequences of ECG signals. This is complemented by a Transformer-based module, which enhances the model's capacity for data-parallel processing, enabling it to capture both local and long-range dependencies within the ECG signals. This hybrid approach significantly boosts the model's ability to detect OSA with high precision. The ExFormer model achieved an accuracy of 85.47%, sensitivity of 79.28%, specificity of 89.15%, and an area under the curve (AUC) of 92.87%, indicating robust performance in classifying OSA. In a comparative analysis with existing methods, the ExFormer demonstrated a 0.37% improvement in accuracy over the best-performing alternative model, signifying its refined capability to detect apnea events. This improvement, though modest, highlights the model's precision and reliability in OSA diagnosis. These findings emphasize the ExFormer model's superior performance, offering an accurate, non-invasive, and convenient tool for clinical applications in OSA detection, which can ultimately lead to more effective patient care and management.