Abstract
BACKGROUND: Electrocardiogram (ECG) signals contain cardiopulmonary information that can facilitate sleep apnea detection. Traditional methods rely on extracting numerous ECG features, which is labor-intensive and computationally cumbersome. METHODS: To reduce feature complexity and enhance detection accuracy, we propose a spectral feature-based approach using single-lead ECG signals. First, the ECG signal is preprocessed via ensemble empirical mode decomposition combined with independent component analysis (EEMD-ICA) to identify the most representative intrinsic mode function (IMF) based on the maximum instantaneous frequency in the frequency domain. Next, Hilbert transform-based time-frequency analysis is applied to derive the component's 2D time-frequency spectrum. Finally, three spectral features-maximum instantaneous frequency (f(emax)), instantaneous frequency amplitude (V), and marginal spectrum energy (S)-are quantitatively compared between normal and sleep apnea populations using an independent-sample t-test. These features are classified via a random forest machine learning model. RESULTS: The f(emax) and IMF7 components of the reconstructed signal exhibited statistically significant differences (p < 0.001) between normal and sleep apnea subjects. The random forest classifier achieved optimal performance, with 92.9% accuracy, 86.6% specificity, and 100% sensitivity. CONCLUSION: This study demonstrates that spectral features derived from single-lead ECG signals, combined with EEMD-ICA and time-frequency analysis, offer an efficient and accurate method for sleep apnea detection.