Performance Analysis of Fuzzy Rough Assisted Classification and Segmentation of Paper ECG Using Mutual Information and Dependency Metric

The paper aims at the development of fuzzy rough set-based dimensionality reduction for discrimination of electrocardiogram into six classes. ECG acquired by the offline method is in the form of coloured strips. Morphological features are estimated using eigenvalues of Hessian matrix in order to enhance the characteristic points, which are seen as peaks in ECG images. Binarization of the image is carried out using a threshold that maximizes entropy for appropriate extraction of the fiducial features from the background. Various image processing algorithms enhance the image which is utilized for feature extraction. The dataset produced comprises the feature vector consisting of 79 features and 1 decision class for 6 classes of ECG. Extensive analysis of dimensionality reduction has been done to have relevant and nonredundant attributes. Fuzzy rough domain has been explored to take into account the extreme variability and vagueness in the ECG. Optimal feature set is subjected to fuzzification using Gaussian membership function. Further, fuzzy rough set concepts help in defining a consistent rule set to obtain the appropriate decision class. Classification accuracy of unfuzzified dataset is compared with the fuzzified dataset. Semantics of the data are well preserved using fuzzy rough sets and are seen from the performance metrics like accuracy, sensitivity and specificity. The proposed model is named as Fuzzy Rough ECG Image Classifier (FREIC) which can be deployed easily for clinical use as well as experimental use.