Stockwell transform empowered attention-guided residual CNN for sleep Apnea classification

Obstructive sleep apnea (OSA) is a chronic sleep disorder linked to severe health conditions such as hypertension and stroke. OSA is typically diagnosed through polysomnography (PSG), an expensive and time-consuming process. To address these limitations, this study proposes an efficient method for OSA detection using a novel time-frequency analysis approach. Traditional techniques for analyzing ECG signals to detect OSA have involved spectrogram and scalogram methods, using Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT) respectively. However, these methods suffer from limitations like spectral leakage and poor time-frequency resolution. To overcome these challenges, we employ the Stockwell transform for feature extraction and segmentation. This transformed data is then fed into a 2D-CNN deep learning model enhanced with channel-wise attention, residual connections, and depth concatenation. Our proposed method demonstrates superior performance, achieving an average accuracy of 95.55%, specificity of 93.64%, sensitivity of 95.55%, and recall of 96.77%. The results show that our framework significantly outperforms existing state-of-the-art methodologies for OSA detection, providing a more efficient and reliable alternative to conventional diagnostic techniques.