Lightweight Arrhythmia Classifier Using Hybrid Compressed Convolutional Neural Network

Convolutional neural networks (CNNs) have been widely applied in arrhythmia diagnosis and electrocardiogram (ECG) classification due to their high accuracy and performance stability. However, memory and computation requirements restrict their wider applications in resource-constrained embedded devices. Also, many existing algorithms simplify networks based on pruning or quantization, directly compresses the size that could not maintain the accuracy with high compression rate. In this paper, we propose a new hybrid compression method based on greedy algorithm to improve the balance compression rate and accuracy. For this purpose, we particularly introduce an evaluation function that can automatically adapt network to its optimal structure. For 17class classification problem on MIT-BIH arrhythmia dataset, the method achieves a $58 \times $ compression rate and suffers only 1.03% accuracy loss, which reduces the storage by $5.6 \times $ as compared to the state-of-the-art network compression structure in ECG classification (from 20.31KB to 3.61KB). Our study further achieves a cost-effective CNN model with high accuracy, and the comprehensive experiments on MIT-BIH arrhythmia dataset also demonstrate the efficiency of the proposed method. Thus, the lightweight convolutional neural network classifier is easier to deploy in embedded devices with limited energy and memory.