Generalizable Beat-by-Beat Arrhythmia Detection by Using Weakly Supervised Deep Learning

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Beat-by-beat arrhythmia detection in ambulatory ECG monitoring is crucial but challenging due to the demanding nature of manual analysis and limitations in current automated methods.
• Existing automatic arrhythmia detection systems struggle with generalization due to insufficient large-sample, finely-annotated ECG data for training.
• The lack of detailed beat-level annotations in large ECG datasets hinders the development of robust and widely applicable arrhythmia detection models.

Purpose of the Study:

• To develop a weakly supervised deep learning framework for arrhythmia detection (WSDL-AD) that enables fine-grained, beat-by-beat analysis using coarsely annotated ECG data.
• To improve the generalization ability of arrhythmia detection models by leveraging large datasets with less granular labels.
• To enhance the accuracy and stability of heartbeat classification under weak supervision through novel techniques.

Main Methods:

• Proposed a weakly supervised deep learning framework (WSDL-AD) integrating heartbeat and recording classification for end-to-end training with only recording-level labels.
• Employed techniques such as knowledge-based features, masked aggregation, and supervised pre-training to enhance weak supervision for heartbeat classification.
• Trained the WSDL-AD model for detecting ventricular ectopic beats (VEB) and supraventricular ectopic beats (SVEB) on multiple large-sample, coarsely annotated datasets.

Main Results:

• The WSDL-AD model demonstrated significant improvements in detection accuracy compared to state-of-the-art supervised learning methods.
• Achieved an 8%-290% improvement in F1 score for supraventricular ectopic beats detection and a 4%-11% improvement for ventricular ectopic beats detection.
• Validated performance on three independent benchmarks according to AAMI recommendations, confirming enhanced generalization and fine detection granularity.

Conclusions:

• The WSDL-AD framework effectively utilizes abundant coarsely labeled ECG data to achieve superior generalization ability compared to previous methods.
• The proposed approach retains fine detection granularity, making it suitable for clinical and telehealth applications.
• This weakly supervised method offers a promising solution for improving the efficiency and accuracy of ambulatory ECG analysis for cardiac arrhythmias.