Sleep Apnea Detection Based on Multi-Scale Residual Network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Traditional convolutional neural networks struggle with complex signal feature extraction for sleep apnea (SA).
• Electrocardiogram (ECG) derived signals, specifically RR interval and R peak signals, are crucial for SA detection.
• Class imbalance in datasets can negatively impact the sensitivity of SA detection models.

Purpose of the Study:

• To propose an effective sleep apnea detection method using multi-scale residual networks.
• To improve the extraction of sensitive signal features from ECG data for SA.
• To overcome the challenge of class imbalance in SA detection.

Main Methods:

• Utilized RR interval and R peak signals derived from ECG as input.
• Employed a multi-scale residual network architecture for robust feature extraction.
• Incorporated a focal loss function to address class imbalance and focus on difficult samples.
• Introduced a fully connected layer for the final SA classification.

Main Results:

• Achieved an accuracy of 86.0%, sensitivity of 84.1%, and specificity of 87.1% on the Apnea-ECG dataset.
• Demonstrated superior recognition accuracy compared to existing methods.
• Effectively mitigated the issue of low sensitivity often caused by class imbalance.

Conclusions:

• The proposed multi-scale residual network offers a promising approach for accurate and sensitive sleep apnea detection.
• The method demonstrates robustness in handling complex signal features and class imbalance.
• This technique holds potential for improving clinical diagnosis of sleep apnea using ECG signals.