Catheter ablation outcome prediction in persistent atrial fibrillation using weighted principal component analysis

Abstract

Radiofrequency catheter ablation (CA) is increasingly employed to treat persistent atrial fibrillation (AF), yet selection of patients who would positively respond to this therapy is currently a critical problem. Several parameters of the surface 12-lead electrocardiogram (ECG) have been analyzed in previous works to predict AF termination by CA. Nevertheless, they are affected by some limitations, such as manual computation and the examination of a single ECG lead while neglecting contributions from other electrodes. AF spatio-temporal organization has been described on surface ECG by means of the normalized mean square error (NMSE) between consecutive atrial activity (AA) signal segments and their reduced-rank approximations based on principal component analysis (PCA). However, these features do not show to be correlated with CA outcome. In this study, such descriptors are adequately adapted and applied to CA outcome prediction. An NMSE index is put forward, computed over the set of eight linearly independent ECG leads after AA signal rank-1 approximations determined by weighted principal component analysis (WPCA). The final predictor is able to discriminate between successful (70.76 ± 17.74) and failing CA procedures (37.54 ± 20.01) before performing the ablation (p-value = 0.0013, AUC = 0.91). The proposed WPCA-based technique emphasizes the most descriptive components of AF electrophysiology by selectively enhancing contributions coming from the most representative ECG leads. Our investigation confirms that ECG spatial diversity exploitation in this WPCA-based framework not only endows the NMSE index with clinical value in the context of CA outcome prediction, but it also improves classification accuracy and increases robustness to ECG lead selection.

Introduction

Atrial fibrillation (AF) is a sustained cardiac arrhythmia characterized by rapid and disorganized atrial activations inducing a loss of atrial mechanical efficacy. Several theories have been suggested to explain AF electrophysiological mechanisms, so as to put forth a systematic procedural protocol for its treatment. AF activity has been first regarded as the result of interactions between multiple wandering atrial wavelets [1], [2]. On the other hand, it is commonly acknowledged that pulmonary veins (PVs) significantly contribute to AF maintenance and evolution, especially in paroxysmal forms of this disease [3]. In spite of major advances in its treatment, AF remains a significant cause of cardiovascular morbidity and mortality, especially those arising from stroke and heart failure.

Radiofrequency catheter ablation (CA) has become the first-line strategy [4] for the treatment of this disease. However, as the precise pathophysiology of AF dynamics has not been completely clarified yet, it is still questionable whether CA effectively suppresses abnormal rhythm sources, and how it affects heart electrical substrate. Different CA techniques have been developed, yet none of them is widely considered as effective for the treatment of persistent AF. Their performance is still far from satisfactory, and they are less effective than equivalent procedures for paroxysmal AF. Since this cardiac interventional procedure is profoundly influenced by operator's experience and patient's health conditions, results reported by clinical centers are quite disparate and not easily comparable [5], [6], [7]. It follows that its efficacy in terminating AF and avoiding its recurrence is not guaranteed for all patients. This situation explains the increasing tendency to attempt an a priori selection of patients who can undergo CA and experience durable sinus rhythm (SR) restoration. Several parameters extracted from the surface ECG have been proposed as potential predictors of CA outcome [8], [9]. For example, prolongation of atrial fibrillation cycle length (AFCL) can be associated with AF termination by CA [10]. In other studies [8], it has been argued that the higher the amplitude of the fibrillatory waves (f-waves) observed on the surface ECG, the more likely procedural success.

In parallel, another line of research aims at noninvasive measures of AF spatio-temporal complexity, with the underlying assumptions that treatment modalities should be chosen and therapy outcome could be predicted on the basis of these measures. In [11], a noninvasive measure of AF organization is assessed by the normalized mean square error (NMSE) values between the atrial activity (AA) signal and its rank-3 approximations determined by principal component analysis (PCA) in lead V₁ [11]. This argument is supported by the hypothesis of a correlation between AF organization and the number and interactions of atrial wavefronts through the heart substrate. The choice of V₁ is justified by the fact that it presents the maximum atrial-to-ventricular amplitude ratio among all ECG leads [12]. In [13], CA performance was shown to influence AF spatio-temporal organization, and its effect can be quantified by variations in NMSE values.

Nevertheless, such parameters are affected by several shortcomings. In the first place, some classical ECG-based descriptors are manually computed [8], [10], so they are subject to operators’ subjectivity and thus prone to errors. Furthermore, as most of them are measured in only one ECG lead, they do not account for information that may be provided by other electrodes. Indeed, ECG analysis is not always straightforward, and visual inspection does not capture AF features underlying the whole ensemble of leads; hence, the limitations of classical single-lead techniques, which do not fully exploit multilead ECG spatial diversity. However, AF spatio-temporal complexity as defined in [11] has not been shown to correlate with CA outcome.

Our investigation focuses on the potential application of the spatio-temporal organization of AA measured on the standard ECG by the NMSE index as a tool to discriminate between successful and failing CA procedures before applying the therapy. Contributions provided from the eight independent ECG leads are expressed in terms on NMSE between successive segments of the actual AA signal and their rank-1 approximations computed by weighted principal component analysis (WPCA), and they are finally combined in a single parameter capable of predicting long-term CA outcome. Thanks to this decomposition, the spatial variability of the standard ECG is taken into account, and the most significant ECG leads are also automatically enhanced by assigning different weights to data based on their estimated relevance.