LOW POWER RESOURCE CONSTRAINED ECG CLASSIFIER FOR EDGE AI

Abstract

Miniaturized wearable Artificial Intelligence-of-Things (AIoT) devices exhibit the need to be resource and energy-efficient. In this thesis, we introduced an optimized AIoT algorithm that utilizes (i) quantized multilayer perceptron (qMLP) for converting ECG signals to binary images and (ii) binary convolutional neural network (bCNN) for ECG classification. The bCNN consists of a fused binary convolution to pooling (bCONP) algorithm and cascaded by two binary dense (bDense) networks utilizing a codesigned function-merging and block-reuse techniques for complexity reduction. We deploy our model into a low-power resource-constrained field-programmable gate array (FPGA) fabric. The model requires 5.8× lesser multiply-and-accumulate (MAC) operations than known wearable CNN models. Our model also achieves a classification accuracy of 98.5%, sensitivity of 85.4%, specificity of 99.5%, precision of 93.3%, and F1-score of 89.2%, while consuming dynamic power of only 34.9 µW. Furthermore, our chosen FPGA is the most compact, about 30.8× smaller, which is appealing for wearable devices.