CNN-based Two Step R Peak Detection Method: Combining Segmentation and Regression.

Area of Science:

• Biomedical Signal Processing
• Machine Learning in Healthcare
• Cardiovascular Monitoring

Background:

• U-Net is effective for semantic segmentation in computer vision and biomedical signal processing.
• Direct U-Net application for R-peak detection in ECG signals presents limitations, often predicting multiple high-probability peaks.
• This necessitates post-processing to identify a unique R-peak location within each QRS complex.

Purpose of the Study:

• To develop an improved R-peak detection method for ECG signals.
• To overcome the limitations of U-Net's pixel-wise classification for precise R-peak localization.
• To ensure a unique peak prediction for each QRS complex using a regression approach.

Main Methods:

• Utilized a regression process for R-peak detection instead of traditional pixel-wise classification.
• Trained the proposed model on diverse ECG data sources, including resting ECG systems, wearable devices, and public databases.
• Investigated the model's robustness with input data from heterogeneous devices, particularly wearable ECG devices.

Main Results:

• The regression-based approach successfully guarantees a unique R-peak location prediction.
• The model demonstrated robustness when trained on various data source combinations.
• Performance was validated across different ECG acquisition methods, including wearable technology.

Conclusions:

• A regression-based method offers a more robust and accurate solution for R-peak detection in ECG signals compared to U-Net's classification approach.
• The proposed model effectively handles data from heterogeneous sources, showing promise for real-world applications.
• This technique enhances the reliability of automated ECG analysis, particularly in scenarios involving wearable monitoring devices.