Low-power, Energy-efficient, Cardiologist-level Atrial Fibrillation Detection for Wearable Devices

2025-08-13

AI Summary

Long-term, continuous atrial fibrillation (AF) monitoring on wearable devices faces a fundamental trade-off between ultra-low power consumption and high diagnostic accuracy. Method: To address this, we propose a hardware-software co-designed neural architecture search (NAS) framework optimized for FPGA-based edge inference. Our approach jointly optimizes neural network topology, fixed-point quantization schemes, and custom hardware accelerator architecture. Contribution/Results: Implemented on a patch-type wearable platform, the system achieves an ultra-low power consumption of 3.8 mW—enabling over 21 days of continuous operation on a single charge—while attaining 95% AF detection accuracy. This represents a 1–3 order-of-magnitude improvement over prior methods and matches the diagnostic performance of cardiologists. The solution significantly enhances the accessibility, sustainability, and clinical deployability of large-scale AF screening.

Technology Category

• Machine Learning: Hardware-aware ML
• Search and Optimization: Algorithm Configuration
• Cognitive Modeling & Cognitive Systems: Affective Computing

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Personalized, context-aware and across-device search
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Atrial fibrillation (AF) is a common arrhythmia and major risk factor for cardiovascular complications. While commercially available devices and supporting Artificial Intelligence (AI) algorithms exist for reliable detection of AF, the scaling of this technology to the amount of people who need this diagnosis is still a major challenge. This paper presents a novel wearable device, designed specifically for the early and reliable detection of AF. We present an FPGA-based patch-style wearable monitor with embedded deep learning-based AF detection. Operating with 3.8mW system power, which is 1-3 orders of magnitude lower than the state-of-the-art, the device enables continuous AF detection for over three weeks while achieving 95% accuracy, surpassing cardiologist-level performance. A key innovation is the combination of energy-efficient hardware-software co-design and optimized power management through the application of hardware-aware neural architecture search. This advancement represents a significant step toward scalable, reliable, and sustainable AF monitoring.

Problem

• Detecting atrial fibrillation with low-power wearable devices
• Overcoming energy constraints for continuous AF monitoring
• Achieving cardiologist-level accuracy in AF detection affordably

Innovation

• FPGA-based wearable monitor with embedded deep learning
• Hardware-software co-design with neural architecture search
• Ultra-low power system enabling multi-week continuous monitoring