OSANet: Detection of Obstructive Sleep Apnea Episodes Using a DarkNet-Based Deep Learning Approach and Time-Frequency Spectrogram Features Derived from Single-Lead ECG

An obstructive sleep apnea (OSA) detection algorithm has successfully been developed employing bag-of-features (BoF) extracted from an ECG spectrogram alongside a deep learning (DL) framework, yielding a commendable classification accuracy and exceptional temporal resolution. Validation of this study utilized overnight ECG recordings from 33 subjects sourced from the Physionet Apnea-ECG database (PAED), exhibiting an average apnea–hypopnea index (AHI) of 30.23 (/h). The investigation focused on determining the optimal spectrogram window duration, examining 60-s intervals to achieve the desired accuracy level. Moreover, specific frequency bands (0.1–50, 8–50, 0.8–10, and 0–0.8 Hz) were isolated to generate BoF, identifying the most suitable frequency band for precise OSA detection. Employing the DarkNet-based CNN classification model, named OSANet, the algorithm was trained and validated to discern AH and non-AH events. The 5-fold cross-validation (5fold-CV) results demonstrated impressive accuracies of 91.9% for the 0.8–10 Hz frequency bands using BoF derived from ECG spectrograms with 60-s time windows.