Analysis of multidomain features for ECG classification

In this work we studied the classification performance of models based on intervals, angles and amplitudes. These features were extracted from both ECG leads and different scales of the wavelet decomposition. The MIT-BIH Arrhythmia database was used, following AAMI recommendations for class labeling and results presentation. The training and testing set and any cross-validation division of the database was made patient-oriented. A floating feature selection algorithm was used to obtain best performing models in the training set. This model was evaluated in the test set obtaining a global accuracy of 90%; for normal beats, sensitivity (Se) 92%, positive predictive value (+P) 85%; for supraventricular beats, Se 88%, +P 93%; for ventricular beats Se 90%, +P 92%. This classifier model based on multidomain features performs better than other state of the art methods, with a fraction of the features.