Lightweight Deep Learning for Atrial Fibrillation Detection: Efficient Models for Wearable Devices

Abstract

The timely and accurate detection of atrial fibrillation (AF) is crucial for an early intervention and for reducing the associated health risks. Wearable technology has emerged as a viable solution for continuous AF monitoring, but deploying accurate AF detection models on resource-constrained devices remains a challenge due to the high computational and memory demands. This study proposes a lightweight and efficient deep learning approach for real-time AF diagnosis in portable devices. We designed a series of convolutional neural network (CNN) models optimized for high accuracy while maintaining a minimal computational footprint. To further enhance efficiency, we explored deep learning compression techniques, including pruning, quantization, and knowledge distillation. Our results demonstrate that the proposed models achieve state-of-the-art accuracy while significantly reducing memory usage and computational complexity, making them suitable for real-time deployment. Additionally, we validated their feasibility by implementing them on a microcontroller, showcasing their practicality for wearable applications. This research paves the way for accessible, low-power, and high-accuracy AF detection in real-world settings, enabling early diagnosis and timely medical intervention without the need for continuous clinical supervision.