A study of complex network features for electrocardiograms and its Applications in atrial fibrillation recognition

Abstract

Atrial fibrillation (AF) is one of the most common arrhythmias in clinics. The traditional diagnosis of AF mainly depends on physicians’ visual observation of electrocardiograms (ECGs), which is an inefficient, time-consuming, and laborious task. Rapidly evolving complex network principles have opened up a new avenue for studying AF rhythm recognition. This paper thoroughly analyzes seventeen existing network features and proposes three novel network features: local efficiency distribution entropy (E_DE), clustering coefficient distribution entropy (C_DE), and degree distribution entropy (D_DE). Different from the existing local efficiency entropy and clustering coefficient entropy, the three distribution entropy features can reflect probability distributions of network features. This paper compares E_DE, C_DE, and D_DE with existing network features by using T-test, box plots, and machine learning models to validate their effectiveness in AF rhythm recognition. The experiments on PhysioNet/CinC Challenge 2017 show that E_DE, C_DE, and D_DE are superior to existing network features and the accuracy of AF recognition can achieve 94.96%, 94.72% and 95.58%, respectively, when using time-domain features plus E_DE, C_DE, and D_DE.