1. Introduction
Paroxysmal arrhythmias are frequently missed by short-duration clinical ECGs, motivating wearable patches capable of continuous multi-day monitoring, which in turn requires that classification happen on-device to avoid the power cost of streaming raw waveform data.
2. Methodology
A 1D CNN with three convolutional blocks was trained on the MIT-BIH Arrhythmia Database to classify heartbeats into five AAMI-defined superclasses, then compressed via post-training INT8 quantisation and deployed on an ARM Cortex-M4 microcontroller integrated into a single-lead adhesive ECG patch with a 100mAh battery.
3. Results
The quantised model retained 98.1 percent classification accuracy on the MIT-BIH test partition, a 0.4 percentage point drop from the full-precision model, while reducing model size from 612KB to 82KB. Bench testing of the prototype patch showed continuous operation for 6.7 days between charges under typical duty cycling.
4. Conclusion
On-device quantised CNN classification enables multi-day wearable arrhythmia monitoring with clinically relevant accuracy. Future work includes a prospective clinical validation study against Holter monitor ground truth.
References
[1] Moody G. B. and Mark R. G., The MIT-BIH Arrhythmia Database, IEEE Engineering in Medicine and Biology, 2001. [2] Hannun A. Y. et al., Cardiologist-level arrhythmia detection with deep neural networks, Nature Medicine, 2019. [3] Jacob B. et al., Quantization and training of neural networks for efficient integer-arithmetic-only inference, CVPR, 2018.