An Advanced Arrhythmia Recognition Methodology Based on R-waves Time-Series Derivatives and Benchmarking Machine-Learning Algorithms

In this paper, we propose an automated decision-making approach to improve the efficiency of arrhythmia recognition. In particular, we focus on recognizing Normal Sinus Rhythms (NSR) from Abnormal Heart Rhythms (AHR). AHR include atrial fibrillation, atrial utter, sinus bradycardia, and supraventricular tachyarrhythmia. Arrhythmia recognition approaches generally involve a feature extraction step designed to describe the heart rhythms and lead the decision-making process. Indeed, we develop an improved feature extraction strategy employing five dynamic patterns, de fined as R-R intervals time series, and its first four absolute derivatives. The R-R intervals refer to the time interval separating two successive R-waves. Therefore, to describe each dynamic pattern, we use 13 feature measures. These measures comprise four time-domain features, six geometric features, and three non-linear features. As a result, a set of 65-features is built and evaluated to determine the most appropriate consistent combination of features. First, we implement a univariate statistical-based feature selection to remove irrelevant features. Then, we construct a model evaluation and selection process composed of dimensionality reduction strategies and machine learning algorithms. The latter serves to de ne the most suitable model based on its ability to discriminate between NSR and AHR. The findings underscore the benefits of this proposed approach, which could serve as valuable decision-making support in the detection of arrhythmias.