Classification of cardiac arrhythmias based on alphabet entropy of heart rate variability time series

Abstract

Abstract Background Symbolic dynamics’ methods provide a description of time series variability that allows inference of new predictive markers. Classification of disorders using symbolic dynamics is accomplished through the use of nonlinear entropies, measured upon encoded series. Method This work applies a recently developed symbolic dynamics method, alphabet entropy (AlphEn) to heart rate variability (HRV) analysis in order to improve automatic classification of cardiac arrhythmias. Experiments are conducted on PhysioNet MIT-BIH Arrhythmia Database. The approach is experimentally compared with other HRV linear and nonlinear feature combinations established in literature. AlphEn is experimentally compared with other common nonlinear entropies: Shannon’s entropy, approximate entropy, sample entropy, etc. Feature selection using symmetrical uncertainty is used for discovering relevant AlphEn features and random forest algorithm is used for arrhythmia classification. Results The best classification result obtained for six heart rhythms on 20 s segments is achieved for AlphEn no-change threshold θ = 100 ms. AlphEn features improved mean sensitivity of other feature combinations by 2% on average, with the best results achieved: SENS: 91.2%, SPEC: 97.1%, AUC: 99.0%. AlphEn may be used efficiently by adding top 10 ranking features, obtained with symmetrical uncertainty, to other established combinations. AlphEn provides the best incremental result to linear feature combination with respect to the inspected entropies. Conclusions AlphEn improves the results of established HRV feature combinations on the problem of automatic cardiac arrhythmia classification. The method enables the extraction of a number of potentially significant, domain-oriented features. It can be used as an accurate first-hand screening for arrhythmia problems.