Deep Learning End-to-End Approach for Precise QRS Complex Delineation Using Temporal Region-Based Convolutional Neural Networks

Deep Learning End-to-End Approach for Precise QRS Complex Delineation Using Temporal Region-Based Convolutional Neural Networks

Paper in proceedings (conference paper)

Advancements in clinical diagnosis of heart disease are driven by technological innovations and signal processing developments. ECG segmentation, particularly QRS complex detection, plays a crucial role in cardiac cycle analysis. Deep learning has revolutionized automated ECG analysis, enhancing diagnostic accuracy significantly. This paper proposes an optimized Region Proposal Network (RPN) architecture for QRS complex detection, specifically designed for 1D signals. Leveraging a vast ECGdataset, extensive data augmentation, feature extraction, and RPN-based QRS detection were employed. Our method achieved a QRS detection F1 score of up to 99 %, highlighting its high reliability. Furthermore, QRS complex delineation exhibited deviations typically within 8 ms. The results were verified using a publicly available Lobachevsky University Electrocardiography Database (LUDB), with validation yielding F1 score of 91.59 % for QRS detection and RMSE of 10.38 ms for QRS complex delineation. The optimized RPN architecture for QRS complex detection presents a promising solution for efficient ECG analysis.

Advancements in clinical diagnosis of heart disease are driven by technological innovations and signal processing developments. ECG segmentation, particularly QRS complex detection, plays a crucial role in cardiac cycle analysis. Deep learning has revolutionized automated ECG analysis, enhancing diagnostic accuracy significantly. This paper proposes an optimized Region Proposal Network (RPN) architecture for QRS complex detection, specifically designed for 1D signals. Leveraging a vast ECGdataset, extensive data augmentation, feature extraction, and RPN-based QRS detection were employed. Our method achieved a QRS detection F1 score of up to 99 %, highlighting its high reliability. Furthermore, QRS complex delineation exhibited deviations typically within 8 ms. The results were verified using a publicly available Lobachevsky University Electrocardiography Database (LUDB), with validation yielding F1 score of 91.59 % for QRS detection and RMSE of 10.38 ms for QRS complex delineation. The optimized RPN architecture for QRS complex detection presents a promising solution for efficient ECG analysis.

ECG Segmentation, Region Proposal Network, QRS Complex, Deep Learning, 1D U-Net

ECG Segmentation, Region Proposal Network, QRS Complex, Deep Learning, 1D U-Net

ŘEDINA, R.; HEJČ, J.; THEURL, F.; NOVOTNÝ, T.; ANDRŠOVÁ, I.; HNÁTKOVÁ, K.; STÁREK, Z.; FILIPENSKÁ, M.; BAUER, A.; MALÍK, M.

12.09.2024

Computing in Cardioligy 2025

Karlsruhe

Computing in Cardiology 2025

2325-887X

Computing in Cardiology

United States of America

1

4

4

https://cinc.org/archives/2024/pdf/CinC2024-107.pdf