This study employed ECG data from the PhysioNet Apnea-ECG Database (Penzel et al., 2000), which contains 70 single-lead recordings classified into three groups based on apnea duration: Group A (≥100 min of apnea events, n = 20), Group B (5–99 min, n = 10), and Group C (<5 min, n = 40). All recordings were approximately 8 h in duration with a 100 Hz sampling frequency. Expert annotators labeled each 1-min segment of the ECG signals, with segments containing ≥1 apnea event marked as “A” and normal breathing segments as “N”.

To minimize borderline cases’ influence, we excluded Group B and focused on 120 recordings (60 from Group A and 60 from Group C), comprising approximately 57,600 1-min segments. The segment distribution showed distinct patterns: Group A contained an A:N ratio of ≈1.8:1 (apnea-dominant), while Group C showed ≈1:19.3 (normal-dominant). The dataset was randomly partitioned at the recording level into training (84 recordings, 40,320 segments) and testing sets (36 recordings, 17,280 segments) in a 7:3 ratio, strictly preserving the original apnea-normal proportion characteristics in both subsets.

The acquisition of ECG signals is frequently contaminated by multiple noise sources, which can be categorized as: Low-frequency baseline wander (0–0.5 Hz), Broadband electromyographic interference (5–2000 Hz), Narrowband power line interference (>50 Hz) (Satija et al., 2018).Conventional denoising approaches based on EEMD typically concentrate significant noise energy in the initial IMF components (Dan et al., 2022). While standard practice involves discarding these noisy IMFs, this procedure inevitably eliminates some diagnostically relevant signal components, thereby compromising subsequent feature extraction and analysis.

To address this limitation, we employ Independent Component Analysis (ICA), which exploits the statistical independence and non-Gaussian characteristics of source signals to achieve superior noise separation (Ramkumar et al., 2021). Our hybrid EEMD-ICA approach demonstrates three key advantages, as illustrated in Figure 2: Effective suppression of multiple noise sources while preserving signal morphology; Significant reduction of baseline wander artifacts; Enhanced signal quality for subsequent HHT processing.

HHT transform is a Time–frequency analysis method with strong adaptability (Zheng et al., 2021). It can obtain the local and global frequency components of the signal based on the non-stationary and nonlinear characteristics of the signal itself, and determine the relationship between the time-frequency energy of the signal. The HHT transform consists of two parts: Empirical Mode Decomposition (EMD) decomposition and Hilbert transform. To solve the problem of modal aliasing during the EMD decomposition process, the Set EEMD is chosen instead of EMD (Dai et al., 2021). Firstly, the EEMD decomposition algorithm is performed as follows:

(1) Add i-fold Gaussian white noise v i n to the original noisy signal x n to form a new signal x i n to be processed (Equation 1):

Then, select the component IMF with the intrinsic characteristics of ECG signal to conduct Hilbert transform (HT) to obtain the corresponding Spectrogram diagram, also known as Hilbert spectrum (Zhang et al., 2016). The specific Hilbert transform process is as follows:

(4) The definition expression of HHT is given by Equation 3:

In the formula: H P i t is the IMF component after Hilbert transform, P i t represents the intrinsic mode component of the electrocardiogram signal.

(5) Construct the analytic signal (Equation 4) from P_i (t):

The marginal spectrum can reflect the energy distribution of each frequency.

The two-dimensional Spectrogram transformation process of ECG signals in this study is shown in Figure 3.

First, the denoised signal is decomposed to extract the intrinsic mode functions (IMFs) of the original signal. Highly correlated IMFs are selected through Pearson correlation analysis and subsequently reconstructed to obtain the refined signal. Next, the maximum instantaneous phase and frequency range (f_emax) of the original signal is determined by reconstructing its marginal spectrum.

During sleep apnea episodes, significant pathological alterations occur in the frequency and amplitude of cardiac and respiratory activities. By analyzing f_emax local variations associated with ECG signals can be characterized. As illustrated in Figure 4, a comparative analysis of f_emax between the two groups reveals a distinct divergence in the marginal spectrum’s maximum instantaneous phase and frequency. Specifically, healthy individuals exhibit a concentration near 2 Hz, whereas obstructive sleep apnea (OSA) patients display a predominant frequency around 4 Hz.

Through comparative analysis of the f_emax ranges across all intrinsic mode functions and the reconstructed signal, as presented in Table 1, we observed that the most prominent alterations associated with sleep apnea pathophysiology were concentrated in IMF5 through IMF7. We then generated Hilbert-Huang transform time-frequency representations for each component by applying the Hilbert transform to individual intrinsic mode functions.

All signal processing and machine learning implementations were performed in MATLAB R2024a (MathWorks, Inc). Specifically, we utilized the Signal Processing Toolbox for EEMD and Hilbert transform computations, and the Machine Learning Toolbox for classification tasks.

Figure 5 presents a representative 30-s electrocardiogram signal comparison between normal sleep and sleep apnea conditions. The Hilbert-Huang transform time-frequency distributions of IMF5 through IMF7 components are displayed, with the upper panel showing results from healthy subjects and the lower panel depicting OSA patients. Color intensity corresponds to signal energy concentration at specific time points.

Analysis of energy distribution across frequency bands reveals that while IMF5 and IMF6 components fall within the reconstructed signal’s maximum instantaneous phase and frequency range, their energy distributions exhibit considerable similarity. These components retain some high-frequency noise, making quantitative feature extraction challenging.

In contrast, IMF7 demonstrates the most pronounced intergroup differences. OSA patients show significantly stronger frequency energy concentrations compared to healthy subjects at corresponding time points. This observation suggests that IMF7 captures sleep apnea-related frequency variations, consistent with known physiological effects of hypoxia and sleep disruption on cardiac activity and ECG signal characteristics (Arnaud et al., 2020).

To further quantify the differences in two-dimensional time-frequency maps between healthy individuals and obstructive sleep apnea (OSA) patients, we selected 45 signal sets from healthy subjects in Dataset C and 40 signal sets from apnea patients in Dataset A. The marginal spectra of the intrinsic mode function (IMF) components IMF5 to IMF7 were computed using the Hilbert transform (HT), as illustrated in Figure 6.

Panels (a), (b), and (c) in the figure display the marginal spectra of IMF5, IMF6, and IMF7 for healthy individuals and OSA patients, respectively. Distinct differences in both the shape and amplitude of the marginal spectra were observed between the two groups. To further investigate these spectral differences, we analyzed the following parameters derived from the marginal spectrum in the frequency domain: Maximum instantaneous frequency, Maximum instantaneous amplitude (V), and Characteristic energy (S) (Wei et al., 2018). The characteristic energy S is defined as (Equation 8): S = ∫ ω 1 ω 2 h 2 ω d ω (8) where w 1 to w 2 represent the natural frequency range of the marginal spectrum, and h ω denotes the Hilbert marginal spectrum.

The differences in three groups of indicators between healthy individuals and OSA patients were analyzed using independent sample t-test (SPSS, Windows version 14.0), and the analysis results are shown in Table 2.

The analysis presented in the table reveals that the quantitative parameters of IMF5 and IMF6 show no statistically significant differences between the two study populations. This observation aligns with the patterns observed in the time-frequency maps. In contrast, the IMF7 component exhibits marked differences in maximum instantaneous phase, frequency amplitude, and characteristic energy between groups. Furthermore, the f_emax of the reconstructed signal demonstrates significant variation, with healthy individuals consistently displaying lower f_emax values compared to patients with obstructive sleep apnea. Patients with OSA conversely show greater amplitude and characteristic energy S values than healthy controls.

This physiological distinction may be explained by the effects of sleep apnea on oxygenation. During apneic episodes, inadequate oxygen supply to the brain and peripheral tissues results in systemic hypoxia, including diminished cardiac oxygenation. In response, the brain initiates compensatory mechanisms such as transient respiratory pulses to elevate breathing frequency, as documented in reference (Hossen and Qasim, 2020). This physiological adaptation consequently increases heart rate, with these pathophysiological changes manifesting clearly in electrocardiogram signals. Through careful examination of the modal and frequency distribution characteristics across ECG signal components, we can effectively differentiate between healthy sleep patterns and those affected by OSA. Of particular note, the IMF7 component combined with the reconstructed signal f_emax emerges as a robust discriminative feature for distinguishing normal sleep from sleep apnea.

To assess the diagnostic potential of these features, we implemented a random forest algorithm with stratified 10-fold cross-validation to evaluate classification performance using three distinct feature combinations: IMF5 quantized features with reconstructed signal f_emax, IMF6 quantized features with reconstructed signal f_emax, and IMF7 quantized features with reconstructed signal f_emax.

As demonstrated in Table 3, this study achieved 92.9% classification accuracy with 86.6% specificity and 100% sensitivity using Random Forest, significantly outperforming conventional methods based on heart rate variability or ECG-derived respiration. The developed spectral features demonstrated consistent performance across multiple machine learning architectures, achieving 87.5% accuracy with SVM and 88.24% accuracy with CNN, while preserving physiological interpretability.

To determine the most discriminating component for OSA detection, we performed a systematic feature ablation study using the IMF7 feature set, the results of which are shown in Table 4. The results reveal IMF7-Characteristic energyS as the most critical feature, with its exclusion causing a substantial 28.19% decrease in classification accuracy. The reconstructed signal’s maximum instantaneous frequency also proved highly influential, showing a 22.31% performance reduction when removed. In comparison, amplitude features demonstrated more moderate importance, with a 16.43% accuracy decline upon exclusion. These findings collectively validate our feature selection methodology, as the complete feature ensemble achieved peak performance at 92.90% accuracy by effectively combining these complementary discriminative characteristics.

The spectral features extracted from ECG signals in this study offer two significant advantages for OSA detection. First, the time-frequency analysis provides clear physiological interpretation of how OSA affects ECG signals while eliminating irrelevant features. Second, the focused approach on key discriminative components such as IMF7 and femax effectively reduces feature dimensionality. This optimization minimizes the negative impact of excessive features on machine learning performance while maintaining excellent adaptability across different classification algorithms.