Performance Evaluation of Multi-Domain Heart Rate Variability Features for Short-Term Arrhythmia Detection in the MIT-BIH Database

Abstract

Heart Rate Variability (HRV) is an index of the autonomic control over the heart and conveys complementary information on cardiac arrhythmias compared to the ECG morphology analyzed by an Automatic Arrhythmia Classification (AAMI) algorithm. Short-term HRV analysis is of interest for real-time assessment and for design of wearable devices. From the MIT-BIH Arrhythmia Database, we derived a set of ECG signals corresponding to samples of Normal Sinus Rhythm (NSR), Ventricular Ectopic Beats (VEB), and Supraventricular Ectopic Beats (SVEB) beats, classified into AAMI-compliant beat superclasses based on derived RR intervals. In each sample of arrhythmia class (episode), we derived two sets of 5 min long subsequent segments, respectively corresponding to the time-, frequency-, non-linear, and time–frequency domains of analysis. We calculated sample entropy (SampEn), Vector Length Index (VLI) and Vector Angle Index (VAI) on each of the derived segments. We carried out the evaluation of the stability of the derived indexes by comparing their average values across the derived pairs of segments (one pair for each of the 5 min long segments composing each episode). We carried out a three-classification problem regarding the discrimination of NSR vs. VEB, NSR vs. SVEB, and VEB vs. SVEB by exploiting 10-fold cross-validation protocol. Results show significant differences of SampEn, VLI and VAI among the three rhythm classes of interest and classification accuracies for the three pairs of classes of 92.0%, 78.8%, and 77.8% for the classifications of NSR vs. SVEB, NSR vs. VEB, and VEB vs. SVEB, respectively. The evaluation of the stability of the indexes across consecutive 5 min long segments gives evidence that the derived indexes vary in a very limited way among the corresponding segments of each episode. This confirms the robustness of short-term HRV analysis in deriving indexes able to represent the autonomic heart control in presence of cardiac arrhythmias, and that, in this context, non-linear and time–frequency derived indexes, as SampEn, VLI, and VAI provide relevant information for automatic arrhythmia classification within an AAMI compliant framework.