Detection of Paroxysmal Atrial Fibrillation by Machine Learning of Lorenz Plot Images

Abstract

[Aims] Atrial fibrillation (AF) is known as a major cause of acute cerebral infarction. Cerebral infarction with AF could be prevented by taking oral anticoagulants. Thus, it is critically important to detect and diagnose the presence of AF. Machine learning of Lorenz plot (LP) images is recognized as a promising method for the detection of AF in long-term ECG monitoring, however, the optimal segment length of LP image is unknown. We examined the performance of AF detection by differing LP segment length using convolutional neural network (CNN).

[Methods] The datasets of 32x32-low-resolution LP images of R-R interval segments between 20 and 600 beats were created from 24-h ECG data in 52 patients with AF and 58 non-AF controls as teacher data and in 53 patients with paroxysmal AF and 52 non-AF controls as verification data.

[Results] In the verification, the positive likelihood ratio (LR) for detecting AF showed a convex parabolic curve with a peak positive LR at 100 beats, while negative LR decreased monotonously as the segment length decreased.

[Conclusions] The optimal segment length using CNN models should be 100 beats for discriminating between AF and non-AF.