Improved abnormality detection from raw ECG signals using feature enhancement

This research presents an abnormal beat detection scheme from lead II Electrocardiogram (ECG) signals along with some improvements on feature extraction. A set of 16 features representing positions, durations, amplitudes and shapes of P, Q, R, S and T waves is proposed in this work for heart beat classification. These features carry important medical information for normal and abnormal beat detection. Diverse classifiers are employed for abnormality detection, including K-Nearest Neighbor, Decision Tree, Artificial Neural Network, Naive Bayesian Classifier, Random Forest, and Support Vector Machine along with some ensemble classifiers such as AdaBoostM1 and Bagging. We have evaluated the proposed system on raw one lead signals extracted from MIT-BIH Arrhythmia, QT and European ST-T databases in the Physionet databank. The experiments using this new set of 16 features achieve better performance for the three test databases than our previous system using a subset of these features.