Learned Vs. Hand-Designed Features for Ecg Beat Classification: a Comprehensive Study

Abstract

In this study, in order to find out the best ECG classification performance we realized comparative evaluations among the state-of-the-art classifiers such as Convolutional Neural Networks (CNNs), multi-layer perceptrons (MLPs) and Support Vector Machines (SVMs). Furthermore, we compared the performance of the learned features from the last convolutional layer of trained 1-D CNN classifier against the handcrafted features that are extracted by Principal Component Analysis, Hermite Transform and Dyadic Wavelet Transform. Experimental results over the MIT-BIH arrhythmia benchmark database demonstrate that the single channel (raw ECG data based) shallow 1D CNN classifier over the learned features in general achieves the highest classification accuracy and computational efficiency. Finally, it is observed that the use of the learned features on either SVM or MLP classifiers does not yield any performance improvement.