End-to-End Premature Ventricular Contraction Detection Using Deep Neural Networks.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate cardiac rhythm assessment in Holter monitoring relies on precise identification of heartbeats and ventricular premature contractions (PVCs).
• Traditional Holter analysis methods face challenges in reliably detecting complex arrhythmias.

Purpose of the Study:

• To introduce and evaluate a novel 1D U-Net neural network for enhanced detection of PVCs in Holter recordings.
• To assess the model's performance in QRS complex detection and PVC identification against established methodologies.

Main Methods:

• Utilized Icentia 11k, INCART DB, and a custom dataset for training the 1D U-Net model.
• Validated the model on AHA DB, MIT 11 DB, NST, and a custom real-world dataset, comparing results with traditional Holter analysis.

Main Results:

• The 1D U-Net model achieved near-perfect balanced accuracy for QRS complex detection across all tested databases.
• Balanced accuracy for PVC detection ranged from 0.909 to 0.986, demonstrating robust performance despite some variability.
• While sensitivity varied, the model's balanced accuracy indicated equitable performance in identifying both false positives and negatives.

Conclusions:

• The 1D U-Net architecture is highly effective for QRS complex detection in Holter monitoring.
• Further research and model refinement are necessary to improve PVC detection accuracy, addressing real-world complexities and noise.