Artificial intelligence-based approaches for sleep-related breathing events identification using EEG and ECG signals

Abstract

Purposes: Sleep apnea or hypopnea is a sleep-related breathing disorder characterized by insufficient ventilation during sleep. Sleep apnea is classified into two major forms: obstructive sleep apnea (OSA) and central sleep apnea (CSA). The conventional diagnosis with Polysomnography (PSG) is time-consuming, uncomfortable, and costly in the clinical setting. To address these issues, wearable devices and AI techniques have been developed, utilizing single or multi-modal physiological signals. This study aims to deploy a multi-modal approach by analyzing both EEG and ECG signals derived from home sleep testing devices for OSA/CSA/hypopnea identification. A robust ensemble learning model is proposed to compare with the performance of the deep learning model in event classification.

Methods: EEG and ECG signals from 201 PSG were collected. Non-linear features extracted by wavelet transform methods and machine learning were used to develop a classification algorithm. ECG spectrograms and the deep learning model were also deployed to compare with traditional method. Two classification strategies including 3-class (OSA-hypopnea-normal, OSA-CSA-normal) and 2-class (OSA-hypopnea, OSA-CSA) were also examined.

Results: The highest classification performance was achieved using the combined signal-based model with 98.8% accuracy, 99.1% sensitivity, and 98.5% specificity for classifying OSA and CSA. When compared with the deep learning model, the classification accuracy of the combined signal-based machine learning model was significantly higher in almost all classification strategies.

Conclusion: The findings highlight the effectiveness of combining non-linear features from ECG and EEG signals for classifying various sleep-related breathing events. A proposed machine learning model provides significantly precise classification compared to a deep learning approach, offering improved reliability in-home sleep setting.

Keywords: Central sleep apnea; Deep learning; Hypopnea; Machine learning; Obstructive sleep apnea; Wavelet transform.

Fig. 1

The flowchart of the proposed classification algorithm. The process of sleep-related breathing events classification based on EEG and ECG signals consists of signal preprocessing, feature extraction and selection, and model construction. These signals are decomposed into different frequency sub-bands using wavelet transform technology which is a two-dimensional signal analysis to obtain time and frequency components simultaneously. In response to reducing feature dimensions, a feature selection method using Neighborhood Component Analysis (NCA) is applied to choose the optimal feature set. The obtained features were then utilized as inputs for the classifiers. Finally, we developed two distinct models: the first one used machine learning for classifying features derived from signal decomposition, while the deep learning approach employed direct ECG spectrograms for sleep-related breathing events classification. The models are evaluated with 5-fold cross-validation and compared to their performance. Additionally, two classification approaches include 3-class (OSA-hypopnea-normal; OSA-CSA-normal) and 2-class (OSA-hypopnea; OSA-CSA). SMOTE synthetic minority oversampling technique, DWT discrete wavelet transform, db Daubechies wavelet, HT Hilbert transform, IA instantaneous amplitude, IF instantaneous frequency, WIF weighted instantaneous frequency, WPA wavelet packet analysis, CWT continuous wavelet transform

Fig. 2

Transform of a 30-second ECG epoch to ECG-spectrogram at 0.5–50 Hz

Fig. 3

Performance in NREM and REM stages for 3-class classification using four different models. (A) The results of OSA, hypopnea, and normal classification. (B) The results of OSA, CSA, and normal classification. The bar chart compares the classification performance through EEG, ECG, combined EEG-ECG, and ECG spectrogram using EBT and ResNet50 by calculating the average percentage of accuracy, sensitivity, and specificity from each class in testing sets