Harnessing TinyML for Accurate ECG Beat Detection

In this paper, we propose the use of a lightweight deep learning-based ECG beat detection model optimized for resource-constrained microcontrollers. Accurate R-peak detection is essential for analyzing heart rate variability and diagnosing arrhythmias. Despite achieving high performance, complex deep learning models often require high computational resources, limiting their applicability in real-time and wearable healthcare systems. To address this challenge, our study focuses on developing an efficient model that balances accuracy and computational efficiency. Our private collected ECG dataset is used for training. Because ECG signals are often noisy and vary across individuals, we implemented preprocessing techniques such as Butterworth band-pass filtering, baseline wander removal to enhance data quality. The proposed model combines traditional signal processing with deep learning to achieve robust QRS complex detection. Our model architecture is inspired and developed from the Fire module to balance high detection accuracy with a low number of parameters and inference time, making it suitable for wearable devices and real-time applications. Compared to a baseline approach, our method demonstrates improved detection accuracy in MIT-BIH dataset. Our model achieved 99.69% sensitivity and 99.69% PPV, outperforming [1] (99.47% sensitivity, 99.6% PPV) and [2] (98.92% sensitivity, 99.35% PPV), demonstrating superior accuracy in ECG beat detection. Experimental results demonstrate that our method effectively identifies R -peaks while maintaining efficiency, contributing to scalable and cost-effective continuous heart monitoring.