CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Classification of heart conditions using functional data analysis Authors:  Chibueze Ogbonnaya - University of Nottingham (United Kingdom) [presenting]
Simon Preston - University of Nottingham (United Kingdom)
Andrew Wood - The University of Nottingham (United Kingdom)
Karthik Bharath - University of Nottingham (United Kingdom)
Abstract: A functional data analysis approach to heart defect detection using heart signals recorded by electrocardiograms (ECGs) is proposed. ECGs can be thought of as continuous functions having an amplitude and phase component. Raw heart signals are usually noisy with artifacts such as baseline wander. When comparing two or more ECGs, there are also issues with arbitrary location and scale. To remove these artifacts and deal with the issues encountered when comparing two or more signals, we propose amplitude registration models and give closed form solutions for the amplitude parameters. For some heart conditions which are characterised by amplitude changes, such as high peaks or inverted curves, to classify the subjects, we first perform functional principal component analysis (FPCA) on the registered functions and use the PC scores as predictors in a classifier. When heart conditions correspond to phase changes, we propose to use a parametric family of warping functions to detect these phase differences. Classification is done using the estimated parameters as predictors in a classifier. The predictive accuracy of our method using leave-one-out cross-validation are 93\% and 100\% for the amplitude classification of Myocardial Infarction and Cardiomyopathy respectively. This compares favourably with existing approaches for classification of ECGs.