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.