An Application Of Mutual Information For Electrocradiogram Feature Selection

Abstract

Correct classification of arrhythmias by implantable cardio-defibrillators (ICDs) and automatic external defibrillators (AEDs) depend on the quality of the features used during classification. Mutual Information provides a means of quantifying and assessing the quality of cardiac related features. This paper proposes a novel algorithm called Weighted Mutual Information or WMI based on mutual information as a method for selecting cardiac related features for classification. To evaluate the algorithm, a comparative analysis was conducted against the Principal Component Analysis (PCA) algorithm using two separate feature sets extracted from the MIT-BIH database: QRS based features and morphology based amplitude values (ECG amplitude values). Two k Nearest Neighbor (k-NN) clustering algorithms were trained using features extracted using WMI and PCA to classify four event types: Normal, Ventricular Ectopic, Atrial Ectopic, and Fusion events. The k-NNs trained using WMI produced a lower classification error when compared to the k-NNs trained using PCA with statistical significance (0.01 > p). Lastly, the k-NN classifier trained on QRS based features outperformed the k-NN classifier trained only on morphology based features with a measurable significant difference.