Implementing ultra-lightweight co-inference model in ubiquitous edge device for atrial fibrillation detection

Implementing internet of things technologies in health monitoring systems attracts a lot of attention. Running the model at edge can continuously and in real-time monitor the user’s physiological information, which can be adopted in universal medical care. However, this task is challenging and rarely discussed due to its limited computational capacities and storage resources. This work propose a novel framework for neural network training on resource-constrained embedded systems. Based on feature engineering, the lightweight neural network is implemented on embedded devices, including the training and testing. By losing some certainty, an ultra-lightweight model can be realized through parameter quantization to get further memory saving and reduce the storage burden of the device. The method is implemented and tested on a cheap AVR-based 8-bit micro-controller for atrial fibrillation detection from electrocardiogram (ECG) features. The results prove the feasibility of 100 samples training on-device with test performance comparable to models developed on a computer. The trained lightweight model can be compressed to about 0.3 of the original size with negligible rebuilt performance loss. It can be combined with wearable devices for chronic diseases and long-term monitoring.