Comparing the Performance of Random Forest, SVM and Their Variants for ECG Quality Assessment Combined with Nonlinear Features

Abstract

For evaluating performance of nonlinear features and iterative and non-iterative classification algorithms (i.e. kernel support vector machine (KSVM), random forest (RaF), least squares SVM (LS-SVM) and multi-surface proximal SVM based oblique RaF (ORaF) for ECG quality assessment we compared the four algorithms on 7 feature schemes yielded from 27 linear and nonlinear features including four features derived from a new encoding Lempel–Ziv complexity (ELZC) and the other 26 features. Seven feature schemes include the first scheme consisting of 7 waveform features, the second consisting of 15 waveform and frequency features, the third consisting of 19 waveform, frequency and approximate entropy (ApEn) features, the fourth consisting of 19 waveform, frequency and permutation entropy (PE) features, the fifth consisting of 19 waveform, frequency and ELZC features, the sixth consisting of 23 waveform, frequency, PE and ELZC features, and the last consisting of all 27 features. Up to 1500 mobile ECG recordings from the Physionet/Computing in Cardiology Challenge 2011 were employed in this study. Three indices i.e., sensitivity (Se), specificity (Sp) and accuracy (Acc), were used for evaluating performances of the classifiers on the seven feature schemes, respectively. The experiment results indicated PE and ELZC can help to improve performance of the aforementioned four classifiers for assessing ECG quality. Using all features except ApEn features obtained the best performances for each classifier. For this sixth scheme, the LS-SVM yielded the highest Acc of 92.20% on hidden test data, as well as a relatively high Acc of 93.60% on training data. Compared with the other classifiers, the LS-SVM classifier also demonstrated the superior generalization ability.