Robust R-peak detection in an electrocardiogram with stationary wavelet transformation and separable convolution

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• R-peak detection is crucial for electrocardiogram (ECG) analysis.
• Previous deep learning models often show performance degradation across different databases.
• Robust R-peak detection is needed for reliable ECG interpretation.

Purpose of the Study:

• To develop a novel deep learning model for accurate R-peak detection.
• To achieve high performance in cross-database validation scenarios.
• To improve the generalizability of R-peak detection algorithms.

Main Methods:

• Developed a deep learning model integrating stationary wavelet transform (SWT) and separable convolution.
• Utilized detail coefficients (level 4) from SWT and filtered ECG derivatives as model inputs.
• Employed separable convolution with atrous spatial pyramidal pooling and noise-augmented training data.

Main Results:

• Achieved high F1 scores (0.9985–0.9999) in cross-database validation on MIT-BIH, INCART, and QT databases.
• Demonstrated consistent high performance on additional testing databases (MIT-BIH-ST, European ST-T, TELE).
• Showed improved performance on the MIT-BIH Noise Stress Test (MIT-BIH-NST) database with noise augmentation.

Conclusions:

• The proposed SWT and separable convolution-based model offers superior R-peak detection performance.
• The model exhibits excellent generalizability and robustness across multiple ECG databases.
• This approach advances the reliability of automated ECG analysis through accurate R-peak detection.