Classification of paroxysmal atrial fibrillation using sinus rhythm electrocardiograms using the symmetric projection attractor reconstruction method

TL;DR

This paper introduces a new method using ECG data to detect paroxysmal atrial fibrillation by transforming signals into images and analyzing them with machine learning.

Contribution

The novel use of symmetric projection attractor reconstruction with machine learning to classify PAF from sinus rhythm ECGs is presented.

Findings

Using KNN with SPAR method achieved 81.2% classification accuracy for PAF detection.

Decision tree models showed 72.5% sensitivity, significantly better than 34% from traditional monitoring.

Abstract

Atrial fibrillation is the most commonly encountered cardiac arrhythmia, increasing stroke risk and mortality. Paroxysmal atrial fibrillation (PAF) can be challenging to detect because arrhythmias occur intermittently. We have been able to classify PAF patients from sinus rhythm electrocardiograms (ECG), using a signal processing technique, Symmetric Projector Attractor Reconstruction, which transforms the ECG time-series into a quantifiable two-dimensional image termed an attractor. To optimise this methodology, we investigated the impact of varying parameters within the SPAR method, choice of lead, ECG sampling frequency and machine learning model choice. We determined that using a K nearest neighbours (KNN) model with 125 Hz ECG sampling frequency, using specific features that quantified the density of the attractor, gave a classification accuracy of 81.2%. When using a decision tree model it was found that the sensitivity was 72.5% which shows an obvious improvement over 30-day long term monitoring which has a sensitivity of 34%. The results of this paper present consideration for the application of a new method to the clinically relevant problem of aiding detection of PAF and conjectures as to the reasoning behind these results. Further investigation in larger cohorts is needed to fully elucidate these findings.

Figures