ECG arrhythmia classification using class-modular MLP

Abstract

ECG (Electrocardiography) is a graphical signal of electrical activity recorded from electrodes on the body surface. It is one of the most important biosignal used by cardiologists for diagnostic purposes. In this study, our main objective is automatically recognition of arrhythmic signal abnormalities, which may be a clue for diagnosis. The detection of an abnormality in ECG signals by human is both complex and error-prone. This motivated researchers to study automatic detection of cardiac arrhythmia disorders, using intelligent data analysis techniques. Computer software using machine learning techniques could easily analyze complex ECG signals, transform signals, make some predictions about the presence of arrhythmia, and provide decision-support information to humans. In this study Multilayer Perceptron (MLP), which is a neural network-based machine learning technique and Class-Modularity concept were applied to two ECG datasets for arrhythmia classification. Class-modularity was also used by class-dependent feature selection to obtain robust modules also providing dimensionality reduction. RELIEF was selected as a well-known technique for class-specific feature list creation. One of the datasets is from UCI repository and it was used on similar studies before. A local dataset is created using real-life ECG recordings collected from Turkish patients. These records are digitized and examined by a medical doctor. The performances of learning methods are improved by feature selection (Decision Trees, SVM-RFE) and feature extraction (PCA) dimensionality reduction techniques. As a comparison, Decision Tree and SVM algorithms have been tested on the arrhythmia dataset. Weka and Matlab were used as machine learning tools during the study. According to test results, MLP performs better than decision trees and similar to SVM on both ECG datasets. The class-modular MLP has slightly less performance, while providing several advantages over MLP.