Energy and sparse coding coefficients as sufficient measures for VEBs classification

Abstract

This paper proposes an efficient model strategy based on the use of two features, Energy (En) and sparse coding coefficients (non-zero elements, Nze), from sparse representation (SR) modeling to detect the arrhythmia. First, we use a set of labeled QRS complexes for dictionary learning. Each category of heartbeats generates a dictionary. Then, SR modeling is applied to unlabelled QRS complexes and thus based on the previously learned dictionaries. The making decision to select the correct class of heartbeats depends on the low value of the SR modeling attributes (En and Nze). These features go beyond the six intuitive properties (Robin Hood, Scaling, Rising Tide, Cloning, Bill Gates, and Babies) that a good sparsity measure ought to fulfill, such as Pietra/Gini indexes (PI, GI), Hoyer, etc. We validate the performance of our strategy on records, including VEBs, from the MIT-BIH Arrhythmia and the Petersburg Institute of Cardiological Technics (INCART) databases. The experimental results show an overall accuracy of 95.51%, 92.04%, and 79.19% with MIT-BIH and 97.67%, 86.73%, and 96.88% with INCART. These results correspond to three scenarios: patient-specific, cross-validation, and independent set, unseen data, respectively. Results are better or comparable to those obtained by the complex machine learning-based methods. The most relevant issues in the area of SR for classification (SRC) are mainly related to the measure of sparsity using recognized standards. In this study, we use a strategy based on model selection rather than machine learning (ML). Then, features were sufficient measures of sparsity and robust for VEBs detection.