Atrial Fibrillation Detection on Low-Power Wearables using Knowledge Distillation

The increasing complexity and memory requirements of neural networks have been slowing down the adoption of AI in low-power wearable devices, which impose important restrictions in computational power and memory footprint. These low-power systems are the key to obtain 24/7 monitoring systems necessary for the current personalized healthcare trend since they do not require constant charging. In this work, we apply Knowledge Distillation to our previously published convolutional-recurrent neural network for cardiac arrhythmia detection and classification. We show that the resulting network halves the memory footprint (138 K parameters) and the number of operations (1.84 MOp) compared to the baseline. By using Knowledge Distillation, this network also achieves significantly higher accuracy after quantization (increase in overall F₁ score from 0.779 to 0.828) and is capable of running into a nRF52832 System-on-Chip from Nordic Semiconductors. This promising result lays the groundwork for deployment on resource-constrained embedded platforms such as micro-controllers of the ARM Cortex-M family, thus potentially enabling continuous detection of cardiac arrhythmias in low-power wearable devices.