The need for standardization of time- and frequency-domain analysis of body surface electrocardiograms for assessment of the atrial fibrillation substrate

This editorial refers to ‘Measures of spatiotemporal organization differentiate persistent from long-standing atrial fibrillation’ by L. Uldry et al., on page 1125 Despite some progress in the earlier decades, the current therapy of atrial fibrillation (AF) is still far from being satisfactory. Antiarrhythmic drugs can restore sinus rhythm only during the first few days after onset of the arrhythmia, are unable to effectively prevent recurrence of AF, and their use carries substantial risk of pro-arrhythmia. Catheter or surgical ablation therapy is effective in patients with paroxysmal AF, but its efficacy to cure persistent AF is still under debate. Moreover, AF ablation is afflicted with a number of potentially serious side effects. Preclinical as well as clinical investigations demonstrate that rhythm control therapy is more successful in individuals with low degree of structural remodelling in the atria. Structural heart diseases and AF itself cause cellular hypertrophy, interstitial fibrosis, inflammatory changes, and amyloidosis, which, in turn, lead to progressive electrical uncoupling between muscle bundles and conduction disturbances. Recent direct contact mapping studies in patients with AF have provided evidence that these alterations result in an increased incidence of conduction block and a higher number and smaller size of separate fibrillation waves.1 This enhancement in the complexity of the substrate for AF is regarded as the key mechanism underlying increasing stability of the arrhythmia in structurally remodelled atria.2 Thus, non-invasive tools for the assessment of AF complexity might be of value for better identification of patients in whom sinus rhythm can be restored and successfully maintained. Quantification of the AF substrate by advanced analysis of surface ECGs appears to be a logical step towards non-invasive quantification of the individual degree of electropathological alterations in the atria. The study by Uldry et al3. published in this issue of the Journal presents a novel classification algorithm to discern persistent from long-standing persistent AF, which is based on the spectral content of the surface ECG. The authors computed the spectral envelope of all possible pairs of precordial leads and used this combined frequency information to determine two multidimensional measures of spectral organization: (i) the multidimensional organization index (MOI) and (ii) the multidimensional spectral entropy (MSE). The best performance to classify the patients in persistent or long-standing persistent AF (84.9%) was achieved by applying quadratic discriminant analysis on both MOI (leads V5, V6) and MSE (leads V4, V6). The authors suggest that this classification tool can potentially be applied in a clinical setting to support therapeutic decision-making. Most importantly, the technique might be used to identify patients eligible for rhythm control therapy. Non-invasive AF substrate complexity analysis and classification has attracted much attention the last decade, and a large number of processing and analysis techniques have been proposed, either to assess the complexity of the AF substrate or to predict therapeutic outcome or AF recurrence after treatment. Literature review of non-invasive AF studies (summarized in Table 1) reveals a large diversity of techniques for data acquisition, data analysis, and research purposes. Studies on prediction of outcome based on surface ECG analysis in AF patients Studies on prediction of outcome based on surface ECG analysis in AF patients Non-invasive measurements are acquired in the form of standard 12-lead ECGs,4–8 a Holter recording,9–12 or body surface potential maps (BSPM),13,14 sometimes complemented by anatomical imaging with reverse calculation of local activation times, which is then called electrocardiographic imaging (ECG-I).15 Data pre-processing may include high- and/or low-pass filtering, down-sampling, and ventricular activity detection followed by cancellation or blanking. Analysis methods can be divided into time-based7–9,13–15 and frequency-based methods,5,12 where time-based methods focus on parameters that are computed in the time domain, such as AF cycle length and fibrillation wave amplitude, while frequency-based methods look at the spectral content of the atrial signals, such as dominant frequency and spectral organization. Some approaches combine parameters from the time and frequency domain.4,10,11,16 An advantage of time-based methods is that one can limit the analysis to TQ intervals only, avoiding the QT segments where ventricular activity might interfere with the atrial activity. Frequency-based methods usually require QRST complex cancellation before further analysis can be performed. Another difference between studies is the use of spatial information. Parameters are computed on a single lead4–7,9–12,16 or an ensemble of leads, ranging from pairs to a full array of body surface electrodes.8,13 Body surface potential maps and ECG-I are particularly rich in spatial information. Some studies attempt to identify atrial activation patterns on the body surface14,17 or reconstruct an epicardial activation map from BSPM measurements and the geometry of heart and thorax derived from computed tomography or magnetic resonance imaging (ECG-I).15 Finally, research purposes vary widely among studies. Some studies test the discriminatory power of a classifier between different clinical types of AF (paroxysmal, persistent, and long-standing persistent).9,11,15 For example, Alcaraz et al9,11. successfully distinguished paroxysmal from persistent AF by automatic quantification of ECG sample entropy. Cuculich et al15. also report a significant difference between these two types, but they used a complexity index derived from the number of wavelets and focal sites observed with their ECG-I methodology. This complexity index was not able, however, to separate persistent from long-standing persistent AF. Other studies aim to develop a novel way to quantify the degree of complexity of AF independent from a clinical classification13,14,17 or investigate the relationship between surface measurements and epicardial atrial recordings.16 Bonizzi et al13. demonstrated the ability of a complexity measure based on principal component analysis of a BPSM to quantify both spatial organization and temporal stationarity of AF, whereas Guillem et al14. showed it was possible to detect inter-individual AF substrate differences using activation maps reconstructed from a BSPM. Predicting outcome—in most cases related to a specific treatment—is a third, popular research subject. Several studies investigated the success rate of an ablation procedure6,8 or the success rate of electrical cardioversion and AF recurrences.5,7 Others focused on spontaneous termination of paroxysmal AF.10,12 The studies differ significantly in the selected recordings and the mathematical techniques used for analysis. Nault et al. studied the relationship between f-wave amplitude and success of catheter ablation and found that a larger f-wave amplitude can predict positive ablation outcome.6 Meo et al. demonstrated that incorporating multi-lead information improves prediction of catheter ablation outcome.8 Bollmann et al. looked at fibrillatory rate and its predictive value for electrical cardioversion and AF recurrence,5 whereas Petersson et al. assessed the discriminatory power of sample entropy applied to AF recurrence after electrical cardioversion.7 Finally, Chiarugi et al. used dominant atrial frequency together with heart rate to predict spontaneous termination of paroxysmal AF.12 Nilsson et al. did the same, only with a combination of time–frequency-domain parameters.10 Most, but not all, of these studies demonstrate a reasonable predictive value of non-invasive ECG-based assessment of AF substrate complexity for success of treatment, recurrences of AF or AF termination. However, Petersson et al7. showed that the discriminatory power of sample entropy was not strong enough to predict maintenance of sinus rhythm after electrical cardioversion. Similarly, Bollmann et a5l. could not find a relation between fibrillatory rate and AF recurrence after pharmacological cardioversion. Whether these discrepancies reflect differences in performance of the respective mathematical analysis techniques or differences in the patient populations enrolled in these studies is currently unknown. The study by Uldry et al. clearly advocates a frequency-based method and it incorporates spatial information by considering pairs of precordial leads. The study is the first to successfully distinguish between persistent and long-standing persistent AF with high classification accuracy and an even higher prediction value for long-standing persistent AF. While this study certainly proves the relevance and applicability of the studied frequency-domain parameters MOI and MSE, it also raises the question of how much usage of spatial information from two leads improves the performance of this approach, compared to, for instance, a single lead or multiple (>2) leads. Moreover, the real contribution of this novel classifier to the problem of AF complexity classification will only become clear when the performance is compared to other complexity parameters featured in the literature, such as fibrillation wave amplitude, sample entropy or f-wave frequency. Prompted by the myriad of AF complexity analysis techniques encountered in the literature, we feel that there is need for standardization of AF substrate assessment to be able to adequately address the clinical relevance of ECG-based AF substrate classification. The first essential task is to define the minimal quality of the acquired ECG recordings as well as the methods and intensity of pre-processing that is required before any analysis can be performed. To decide which ECGs are suitable for further analysis, standardized quality criteria for body surface measurements need to be determined. Secondly, the added value of spatial information has to be evaluated to answer the question as to how many leads must be included in an analysis for optimization of outcome prediction. Related to this is the question as to whether BSPMs really enable better substrate assessment than simpler lead configurations, such as a standard 12-lead ECG. Another point of interest is the quantification of the temporal variability of AF and the minimal recording duration necessary to get a stable estimate of a parameter that is used for substrate classification. Only after having defined all these parameters can the relative performance of the different proposed analysis methods be compared and validated in large-scale studies. In the end, the goal is to design, validate, and implement robust ECG-based measures for outcome prediction of rhythm control therapy in the clinic. The study of Uldry et al. not only provides a further and important step in this direction but also contributes to the evidence for the need of standardization of time- and frequency-domain analysis of surface ECGs for AF substrate classification. Conflict of interest: none declared.