Delineation of the electrocardiogram with a mixed-quality-annotations dataset using convolutional neural networks

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Electrocardiogram (ECG) analysis is vital for clinical practice.
• Current digital signal processing (DSP) methods require extensive adaptation for new ECG morphologies.
• Automated ECG interpretation demands accurate detection and delineation of waveform components.

Purpose of the Study:

• To investigate the efficacy of the U-Net deep learning model for ECG wave detection and delineation.
• To evaluate U-Net's performance against traditional DSP algorithms.
• To explore architectural variations and regularization techniques for optimizing U-Net on limited ECG data.

Main Methods:

• Adapted U-Net, an image segmentation deep learning network, for ECG signal processing.
• Trained and validated the model on the PhysioNet QT database using fivefold cross-validation.
• Employed regularization techniques including semi-supervised pre-training, data augmentation, and in-built model regularizers.
• Tested various architectural configurations (depth, width) and inference strategies (single- and multi-lead).

Main Results:

• The best U-Net configuration achieved high precision and recall for P, QRS, and T waves, comparable to DSP methods.
• Precisions: 90.12% (P), 99.14% (QRS), 98.25% (T).
• Recalls: 98.73% (P), 99.94% (QRS), 99.88% (T).

Conclusions:

• Deep learning, specifically U-Net, offers a viable alternative for ECG analysis, even with limited datasets.
• The U-Net approach demonstrates robustness and adaptability for detecting and delineating ECG waves.
• This study highlights the potential of deep learning to overcome limitations of traditional DSP methods in ECG interpretation.