EfficientNet-based machine learning architecture for sleep apnea identification in clinical single-lead ECG signal...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Obstructive sleep apnea (OSA) is a prevalent condition affecting cardiovascular health.
• Accurate and accessible diagnostic tools for OSA are crucial for timely intervention.
• Electrocardiography (ECG) signals offer a potential non-invasive source for OSA detection.

Purpose of the Study:

• To develop a machine learning architecture for identifying OSA patterns in single-lead ECG signals.
• To achieve high performance in detecting OSA using clinical datasets.
• To evaluate the model's effectiveness in a real-world clinical setting.

Main Methods:

• Utilized a dataset of 1656 diverse patients from China Medical University Hospital.
• Employed EfficientNet for apnea segment detection and feature extraction from ECG.
• Compared data preprocessing techniques like overlapping slicing and sample weights.
• Integrated the trained model with XGBoost for patient screening (AHI > 30).

Main Results:

• EfficientNet with overlapping slicing and sample weights achieved an AUC of 0.917 and accuracy of 0.855 for apnea segment detection.
• The combined model with XGBoost yielded an AUC of 0.975 and accuracy of 0.928 for screening patients with severe OSA.
• Performance on PhysioNet data was comparable to existing literature models for OSA screening.

Conclusions:

• The proposed machine learning architecture and optimized training techniques demonstrate strong performance for OSA diagnosis.
• The model's effectiveness with a diverse demographic dataset supports its potential for practical clinical implementation.
• This research advances the non-invasive detection of OSA using ECG, contributing to improved patient care.