Comparing Performance of Iterative and Non-Iterative Classifiers for 2-Lead ECGs on Multi-Feature Schemes

Abstract

In order to evaluate not only performance of iterative classifiers i.e. support vector machine (SVM), least-squares SVM (LS-SVM)) and non-iterative classifier (i.e. random forest (RM), but also efficiency of various feature schemes, the aforementioned machine learning algorithms were applied on six feature schemes of ECG recordings. The ECG recordings were obtained from the MIT-BIH normal sinus rhythm database, the MIT-BIH atrial fibrillation database and the MIT-BIH ST change database. These recordings were initially filtered by a 0.1 - 12 Hz band pass filter. Then 80 features including 48 time domain, 18 frequency domain, 12 time-frequency and 2 principle component analysis (PCA) features were extracted to construct six feature schemes. The SVM, LS-SVM and RF algorithms were used to discern normal, atrial fibrillation (AF) and ST change ECG recordings on the six feature schemes respectively. Finally experiment results illustrated that RF could yield the best highest F1 scores 0.8908 for the binary classification (i.e. normal and AF ECG groups) and 0.7535 for the multi-class classification (i.e. the normal, AF and ST change ECG groups).