Atrial Fibrillation Detection From Electrocardiogram Signal on Low Power Microcontroller

In this paper, we proposed the implementation of a simple and lightweight Atrial Fibrillation detection based on an improved Variable Step Dynamic Threshold Local Binary Pattern algorithm. Using feature selection based on correlation and statistical significance, we can reduce the input feature size to just 44 features without significantly degrading classification accuracies. Tested on 15-second signal segments from the MIT-BIH Atrial Fibrillation Database and combined with a support vector machine classifier, the proposed method can achieve 99.14% sensitivity, 99.12% specificity, and 99.13% accuracy. When the input signal length is 60 seconds, the sensitivity, specificity, and accuracy are 99.49%, 99.46%, and 99.47%, respectively. The reduced input feature size results in a machine learning model size as small as 132.86kB when the input signal length is 60 seconds. When implemented on an Arm Cortex M4-based STM32F413ZHT3 microprocessor with 100MHz clock frequency, the proposed method can achieve similar performance as a PC-based system with an average current consumption of just 27mA. The embedded C program can fit in as small as 114.46kB of flash memory and complete one SVM inference as fast as 11.28ms. The results presented in this paper show that it is possible to do highly accurate machine learning classification that can detect Atrial Fibrillation from ECG signals on a low-power, constrained resource microcontroller. Our results will make developing a high quality, low cost, and low power wearable smart medical electronic device for detecting atrial fibrillation from ECG signal much easier.