AF-DETR: Transformer-Based Object Detection for Precise Atrial Fibrillation Beat Localization in ECG.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Atrial fibrillation (AF) detection in electrocardiograms (ECGs) is challenging at the heartbeat level.
• Current deep learning methods often classify entire ECG segments, missing individual heartbeat anomalies.
• Precise AF detection requires granular analysis of individual heartbeats.

Purpose of the Study:

• To develop a novel deep learning model for precise heartbeat-level AF detection.
• To improve both the localization and classification accuracy of AF heartbeats.
• To validate the model's performance across multiple public ECG datasets.

Main Methods:

• Introduced AF-DETR, a transformer-based object detection model utilizing a CNN backbone.
• Employed a transformer encoder-decoder architecture with 2D bounding boxes for heartbeat representation.
• Implemented contrastive denoising training to enhance convergence and reduce redundant predictions.

Main Results:

• AF-DETR achieved state-of-the-art F1-scores for heartbeat-level classification (up to 99.87%) and segment-level accuracy (up to 99.99%).
• Demonstrated high performance across five diverse public ECG datasets (CPSC2021, AFDB, LTAFDB, MITDB, NSRDB).
• Showcased effective AF heartbeat localization and classification capabilities.

Conclusions:

• AF-DETR significantly improves AF detection accuracy at the individual heartbeat level.
• The model exhibits strong generalization capabilities across various ECG datasets.
• This approach offers a promising solution for precise AF diagnosis using ECG analysis.