Principal component analysis and cluster analysis for measuring the local organisation of human atrial fibrillation

Abstract

The distribution of atrial electrogram types has been proposed to characterise human atrial fibrillation. The aim of this study was to provide computer procedures for evaluating the local organisation of intracardiac recordings during AF as an alternative to off-line manual classification. Principal components analysis (PCA) reduced the data set to a few representative activations, and cluster analysis (CA) measured the average dissimilarity between consecutive activations of an intracardiac signal. The data set consisted of 106 bipolar signals recorded on 11 patients during electrophysiological studies for catheter ablation. Performances of PCA and CA in distinguishing between organised (type I) and disorganised (type II/III, Wells criteria) were assessed, in comparison with manual reading, by evaluating the predictive parameters of the classification analysis. Both methods gave high accuracy (92% for PCA and 89% for CA), confirming the feasibility of on-line characterisation of AF. Sensitivity was lower than specificity (81% against 98% for PCA, and 77% against 97% for CA), with seven out of eight misclassifications of PCA in common with CA. Differences between manual and computer analysis may be related to the higher resolution of PCA and CA in the measurement of the organisation of atrial activations. These procedures are suitable for providing automatic (by CA) or semi-automatic (by PCA) measures of the extent of local organisation of AF in the pre-ablation treatment phase.