Robust Method for Screening Sleep Apnea With Single-Lead ECG Using Deep Residual Network: Evaluation With Open...

Area of Science:

• Cardiology
• Medical Devices
• Artificial Intelligence in Healthcare

Background:

• Sleep apnea syndrome (SAS) diagnosis often requires complex and costly procedures.
• Electrocardiogram (ECG) signals contain physiological information related to respiratory events.
• Developing non-invasive and accessible screening tools for SAS is crucial.

Purpose of the Study:

• To propose and validate a robust method for screening sleep apnea syndrome (SAS) using a single-lead electrocardiogram (ECG).
• To develop models for minute-by-minute abnormal breathing detection and apnea-hypopnea index (AHI) estimation.
• To assess the performance of deep learning models (ResNet18, ResNet34, ResNet50) trained on ECG-derived features.

Main Methods:

• Calculation of heartbeat interval and ECG-derived respiration (EDR) from single-lead ECG.
• Training of ResNet18, ResNet34, and ResNet50 models using heartbeat interval and EDR.
• Development and evaluation using multiple open datasets and experimental data, including wearable device data.

Main Results:

• ResNet18 achieved the highest performance in abnormal breathing detection (Cohen's kappa: 0.57).
• SAS patient classification (AHI threshold 15) yielded an average Cohen's kappa of 0.71.
• Method demonstrated robust performance across diverse datasets, with significant improvement (kappa: 0.91) after tuning on wearable device data.

Conclusions:

• The proposed single-lead ECG-based method offers a robust and accurate approach for SAS screening.
• The method shows equivalent performance to existing studies using open datasets, despite not being trained on them.
• This approach has the potential to reduce development costs for commercial SAS screening software due to its reliance on open datasets and strong generalizability.