Features selection for arrhythmia diagnosis using Relief-F algorithm and support vector machine

Electrocardiography (ECG) is a standard clinical process to record the electrical activity of the heart. It is the most available effective method for diagnosing cardiac arrhythmias. The ECG signals may be classified into either normal or abnormal based on the timing and the potential information of the electrical waves propagating through the heart's muscles. Many advanced signal processing techniques have been used to extract different numerical and logical information from these signals. Usually, the number of extracted features is very high and some of them are redundant, irrelevant and/or noisy. In this work, a thorough experimental study was conducted to reduce the number of ECG features finding a more compact representation of samples selecting the most informative features and removing the others. We empirically investigated the efficiency of two different filter-based feature-selection algorithms for the diagnosis of cardiac arrhythmia: Relief-F and information-gain. We used support vector machine and logistic regression as classification models. Relief-F is a promising filter-based feature selection algorithm. It is a simple and effective algorithm capable of evaluating the feature's importance considering dependence from other features. Classification performance is calculated with three different statistical measures; classification accuracy, sensitivity and specificity. Experimental results showed that the performance of Relief-F with SVM is promising for the diagnosis of cardiac arrhythmia.