LTH-ECG: Lottery Ticket Hypothesis-based Deep Learning Model Compression for Atrial Fibrillation Detection from...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Atrial Fibrillation (AF) is a significant cause of cardiovascular morbidity, including stroke and heart failure.
• Electrocardiograms (ECG) are crucial for diagnosing AF, with single-lead ECG offering practical advantages for portable devices.
• Current deep learning (DL) models for AF detection from ECG are highly accurate but too large for practical deployment on resource-constrained devices.

Purpose of the Study:

• To develop a computationally efficient deep learning model for single-lead ECG-based Atrial Fibrillation detection.
• To significantly reduce the model size of existing DL algorithms without compromising diagnostic performance.
• To enable the deployment of AF screening tools in wearable and implantable devices.

Main Methods:

• Proposed LTH-ECG, a novel goal-driven method based on the Lottery Ticket Hypothesis (LTH) for iterative model pruning.
• Implemented LTH-ECG to achieve over-pruning avoidance during the model compression process.
• Evaluated the performance of the compressed model on single-lead short-time ECG signals for AF classification.

Main Results:

• LTH-ECG achieved a 142x reduction in model size.
• The model compression resulted in an insignificant loss of classification performance, with less than a 1% penalty in test F1-score.
• The compressed model maintains cardiologist-level accuracy for AF detection.

Conclusions:

• LTH-ECG effectively addresses the deployment challenges of DL models for AF detection in wearable devices.
• The method enables practical remote screening of AF, empowering patients with on-demand monitoring capabilities.
• LTH-ECG serves as an early warning system for timely detection and management of atrial fibrillation.