Atrial Fibrillation Detection on the Embedded Edge: Energy-Efficient Inference on a Low-Power Microcontroller.

Atrial Fibrillation (AF) is a common yet often undiagnosed cardiac arrhythmia with serious clinical consequences, including increased risk of stroke, heart failure, and mortality. In this work, we present a novel Embedded Edge system performing real-time AF detection on a low-power Microcontroller Unit (MCU). Rather than relying on full Electrocardiogram (ECG) waveforms or cloud-based analytics, our method extracts Heart Rate Variability (HRV) features from RR-Interval (RRI) and performs classification using a compact Long Short-Term Memory (LSTM) model optimized for embedded deployment. We achieved an overall classification accuracy of 98.46% while maintaining a minimal resource footprint: inference on the target MCU completes in 143 ± 0 ms and consumes 3532 ± 6 μJ per inference. This low power consumption for local inference makes it feasible to strategically keep wireless communication OFF, activating it only to transmit an alert upon AF detection, thereby reinforcing privacy and enabling long-term battery life. Our results demonstrate the feasibility of performing clinically meaningful AF monitoring directly on constrained edge devices, enabling energy-efficient, privacy-preserving, and scalable screening outside traditional clinical settings. This work contributes to the growing field of personalised and decentralised cardiac care, showing that Artificial Intelligence (AI)-driven diagnostics can be both technically practical and clinically relevant when implemented at the edge.

What is it about?

The study focused on developing a system for real-time Atrial Fibrillation (AF) detection using a low-power microcontroller unit (MCU). The methodology involved extracting Heart Rate Variability (HRV) features from RR-Interval (RRI) data and classifying them using a compact Long Short-Term Memory (LSTM) model optimized for embedded deployment. The HRV data was sourced from the MIT-BIH AF Database, comprising ECG recordings annotated by expert cardiologists. A sliding window approach was employed to segment RRI sequences for model development, utilizing a window length of 40 RRIs and a step size of 1. The study achieved an AF classification accuracy of 98.46% with an inference time of 143 ± 0 ms and power consumption of 3532 ± 6 µJ per inference. The results demonstrated that the system is practical for continuous monitoring, balancing classification performance with computational efficiency.

Why is it important?

This study is important as it presents an innovative approach to real-time Atrial Fibrillation (AF) detection using a low-power microcontroller unit (MCU), moving away from reliance on full ECG waveforms or cloud-based systems. By implementing a compact LSTM model optimized for embedded deployment, the research offers a practical and energy-efficient solution for continuous AF monitoring. This development is significant as it addresses the challenge of undiagnosed AF cases, enabling early intervention and reducing the risk of stroke and other adverse outcomes. The approach enhances privacy and extends battery life by minimizing wireless communication, making it suitable for wearable devices outside traditional clinical settings.

Key Takeaways:

1. High Classification Accuracy: The study achieves an overall classification accuracy of 98.46% for AF detection, demonstrating the effectiveness of using HRV features and a compact LSTM model on a resource-constrained MCU.

2. Energy Efficiency: The system performs inferences in 143 ± 0 ms while consuming only 3532 ± 6 µJ per inference, enabling long-term deployment without frequent battery replacements and reducing the need for constant wireless communication.

3. Clinically Meaningful Monitoring: The research supports the feasibility of performing reliable AF monitoring directly on edge devices, offering an energy-efficient and privacy-preserving solution that can be scaled for use in personalized cardiac care outside clinical environments.