WSDL-AD

WSDL-AD performs beat-by-beat detection of cardiac arrhythmias from ambulatory electrocardiogram (ECG) recordings using weakly supervised deep learning to enable training from coarsely annotated, recording-level labels.

Key Features:

• Weak Supervision: Utilizes large amounts of coarsely annotated ECG data where labels are assigned to entire recordings rather than individual beats for training.
• End-to-End Deep Neural Network: Integrates heartbeat classification and recording classification into a unified architecture permitting end-to-end training from recording-level labels.
• Knowledge-Based Features: Incorporates domain-specific, knowledge-based features to inform feature extraction.
• Masked Aggregation: Applies masked aggregation to aggregate information while masking selected data points to focus learning on relevant features.
• Supervised Pre-Training: Employs supervised pre-training strategies prior to fine-tuning with weakly labeled data.

Scientific Applications:

• Supraventricular Ectopic Beat (SVEB) Detection: Enables beat-by-beat detection of SVEBs in ambulatory ECG recordings.
• Ventricular Ectopic Beat (VEB) Detection: Enables beat-by-beat detection of VEBs in ambulatory ECG recordings.
• Cross-Dataset Generalization: Facilitates generalization of arrhythmia detection models across diverse, coarsely annotated ECG datasets for clinical monitoring applications.

Methodology:

Training uses weak supervision with coarsely annotated ECG recordings and end-to-end learning that integrates heartbeat and recording classification; the framework incorporates knowledge-based features, masked aggregation, and supervised pre-training; it was evaluated on five large-sample coarsely annotated ECG datasets and tested on three independent benchmarks recommended by the Association for the Advancement of Medical Instrumentation (AAMI), reporting F1 improvements for SVEB detection of 8%–290% and for VEB detection of 4%–11% compared to state-of-the-art supervised methods.