Compressed Deep Learning Models for Wearable Atrial Fibrillation Detection through Attention.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Deep learning (DL) models show potential for atrial fibrillation (AF) detection using electrocardiogram (ECG) and photoplethysmography (PPG) data.
• Deploying complex DL models on resource-constrained wearable devices presents significant challenges due to computational and memory limitations.

Purpose of the Study:

• To develop a method for compressing DL models for AF detection, enabling their deployment on wearable devices.
• To investigate the effectiveness of a customized channel attention mechanism for model compression and feature selection.
• To enhance the explainability of AF detection models.

Main Methods:

• Integration of a customized channel attention mechanism into DL neural networks for AF detection.
• Application of channel attention for compressing DL models by focusing on salient time-series features.
• Evaluation of model performance on key AF databases (ADB and C2017DB) before and after compression.
• Analysis of learned channel attention distributions to understand model explainability.

Main Results:

• Channel attention significantly reduced the total number of model parameters and file size.
• Minimal loss in AF detection accuracy was observed post-compression.
• Certain compressed model variants showed improved performance on AF databases.
• Analysis of attention distributions highlighted salient temporal ECG/PPG features crucial for AF diagnosis, enhancing model explainability.

Conclusions:

• Integrating attention mechanisms is an effective strategy for compressing DL models for AF detection.
• The proposed approach yields compressed, accurate, and explainable AF detectors suitable for low-power wearable devices.
• Channel attention facilitates the development of simpler, more accurate algorithms for AF screening, offering valuable clinical insights into temporal biomarkers.