A Hybrid Feature Extraction Method to Detect Atrial Fibrillation from Single Lead ECG Recording

Abstract

Identifying patients with Atrial Fibrillation (AFib) is one of the most challenging and prevailing problems in cardiology. In this study, we propose a novel feature extraction method hybridizing probabilistic symbolic pattern recognition (PSPR) and Sample Entropy (SampEn) to represent morphological changes in electrocardiogram (ECG) recordings. We implemented a PSPR framework on continuous SampEn and RR interval series obtained from 4,630 ECG recordings in the training dataset. In our hybrid feature extraction method, PSPR symbolically discretizes SampEn and RR interval series with five and seven unique symbols, respectively and then models the pattern transition behavior of these series using probability theory. We obtained 28 features including PSPR-based metrics and descriptive metrics from SampEn, RR intervals, and processed ECG recordings. A random-forest classifier was trained on 13 features derived using a Genetic algorithm based feature selection technique. On the test dataset of 1,158 ECG recordings, we achieved an accuracy, sensitivity, and specificity of 95.3%, 77.7%, and 97.9%, respectively. Results demonstrate that our proposed hybrid method can extract features that are significant to detect AFib rhythms from single lead short ECG recordings.