QRS detection and classification in Holter ECG data in one inference step.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Wearable devices generate noisy Holter ECG data during daily activities, posing challenges for traditional QRS detection and classification.
• Existing methods often require separate steps for QRS detection and classification, increasing complexity.

Purpose of the Study:

• To develop a deep-learning model for simultaneous QRS complex detection and classification in single-lead Holter ECG.
• To address the challenges of noise and artifacts in wearable ECG data.

Main Methods:

• A deep-learning model comprising ResNet blocks and a gated recurrent unit layer was developed.
• The model was trained and validated on a private dataset of 12,111 Holter ECG recordings and tested on public databases.
• A novel approach integrated QRS detection and classification into a single inference step.

Main Results:

• The model achieved an F1 score of 0.99 for QRS detection on the private test set.
• Cross-database QRS detection performance yielded a mean F1 score of 0.96 ± 0.06.
• QRS classification achieved micro and macro F1 scores of 0.96 and 0.74, respectively, on the private test set.

Conclusions:

• The proposed deep-learning model reliably detects and classifies QRS complexes in a single inference step.
• The method demonstrates superior QRS detection performance compared to existing approaches across multiple public databases.
• This approach offers a robust solution for analyzing noisy Holter ECG data from wearable devices.