1. Introduction
Atrial fibrillation (AF)—the most common sustained cardiac arrhythmia—while not a life-threatening condition itself, leads to an increased risk of stroke or heart failure and high rates of mortality [
1]. Therefore, early detection and diagnosis of AF is a critical issue for all health stakeholders. Identifying individuals at a higher risk of developing AF is feasible using predictive models based on clinical variables, circulating biomarkers, and P-wave indices (PWIs) [
2].
The P-wave on the resting 12-lead electrocardiogram (ECG) is a representation of atrial depolarization, providing a wealth of information valuable in predicting incident AF. The PR interval and other standard PWIs (SPWIs) such as P-wave duration, P-wave terminal force in lead V
1 (PTFV
1), interatrial block (IAB), P-wave area, and P-wave axis, along with QT interval, ECG-derived ventricular hypertrophy, and incidence of atrial and ventricular ectopy have been proposed as potential identifiers of individuals at risk for developing AF [
3]. Similarly, clinical risk scores such as the CHA
2DS
2-VASc score have been directly associated with the incidence of new-onset AF [
4]. Furthermore, the association of PWIs with conditions beyond AF underlines their clinical importance. Therefore, PTFV
1, P-wave duration, advanced IAB (a-IAB), and P-wave area have been correlated with stroke incidence [
5,
6]. Moreover, the addition of the P-wave axis to the standard CHA
2DS
2-VASc score improved its performance in stroke prediction [
7].
More sophisticated ECG indicators derived from P-wave wavelet analysis have been used in patients with no evidence of structural heart disease or any other predisposing factors to successfully identify those at risk of developing AF [
8]. It has also been shown that during sinus rhythm, multiple P-wave morphologies are present more frequently in patients with a history of AF [
9], indicating differences in the electrical substrate of the atria [
10]. Thus, morphological variability, along with beat-to-beat (B2B) wavelet analysis, is proposed as an integrated approach to P-wave analysis.
The present study aims to compare the performance of B2B P-wave analysis with that of traditional predictors of paroxysmal atrial fibrillation (PAF), such as SPWIs, in identifying patients with a history of PAF during sinus rhythm. The superiority of a B2B-P-wave analysis-based classifier over a SPWIs-based one would support the hypothesis that P-wave B2B morphological alterations provide a wealth of information on atrial predisposition to AF that SPWIs cannot reveal.
2. Materials and Methods
We invited 40 consecutive patients with newly diagnosed (less than a month) PAF and no history of antiarrhythmic medication to participate in the study; 60 age- and gender-matched individuals without any history of PAF or structural heart disease were used as the control group. Patients with comorbidities, such as previous cardiovascular surgery, previous cardiac ablation, heart failure NYHA class III-IV, severe valvular heart disease, prosthetic valves, reduced life expectancy, age > 75 years, atrioventricular block, presence of implanted pacemaker or cardiac defibrillator, moderate/severe renal or hepatic impairment, were excluded from the analysis. A complete medical history was obtained from all study volunteers, and both groups underwent clinical examination and calculation of the CHA
2DS
2-VASc score. Eventually, 33 patients with a PAF history and 56 controls fulfilled inclusion criteria. The baseline characteristics of the two groups can be found in
Table 1.
Standard 12-lead ECGs were obtained from all participants. The recordings were performed at least 7 days apart from any AF episode to minimize the potential effect of myocardial stunning. The ECGs were scanned, stored in digital format, magnified sufficiently, and analyzed with digital image processing software (
imagej.nih.gov/ij, accessed on 21 October 2020). Additionally, three-orthogonal axis system (X-frontal, Y-vertical, and Z-sagittal axis) vectorcardiographic (VCG) signals of 10 min duration were also recorded at the same time, with study individuals resting in the supine position, using a high sampling rate (1000 Hz) Galix GBI-3S Holter monitor.
All participants were informed about the scope of the study and gave written informed consent. The study complied with the Declaration of Helsinki and was approved by the Special Purpose General Assembly of Aristotle University School of Medicine (#8, approved on 9 September 2016).
2.1. Standard P-Wave Indices
2.1.1. The 12-Lead ECG Indices
Eight conventional 12 lead ECG PWIs, PR duration, P-wave duration, P-wave dispersion, P-wave peak time, P-wave axis, P-wave voltage in lead I, PTFV
1, and P-wave area were measured by three observers, and mean values were calculated. P-wave dispersion was defined as the difference between the longest, and the shortest P-wave duration measured in any of the standard ECG leads. P-wave peak time is a novel index, equal to the duration between the beginning and peak of the P-wave measured in leads II or V
1 [
11]. An abnormal P-wave axis was determined as a frontal P-wave axis less than 0° or more than 75° [
3], while PTFV
1 was calculated as the amplitude-duration product of the terminal negative component of the P-wave in lead V
1 [
12]. P-wave area (mV × ms) was measured as the sum of the absolute areas underneath the positive and negative P-wave deflections, and the maximum area from among the 12 leads was selected [
3].
Furthermore, P-wave biphasicity in inferior leads was assessed to identify IAB type [
13]. Partial-IAB (p-IAB) was defined as a P-wave ≥ 120 ms without a negative deflection in the inferior leads (II, III, aVF), and advanced IAB (a-IAB) as a P-wave ≥ 120 ms, along with biphasic morphology in inferior leads. Finally, a relatively new composite score, the MVP score (morphology–voltage–P-wave duration), was calculated, assigning up to two points to each of the three components [
14] (
Appendix A.1.1,
Appendix A.1.2,
Appendix A.1.3,
Appendix A.1.4,
Appendix A.1.5,
Appendix A.1.6,
Appendix A.1.7).
2.1.2. Orthogonal Morphology
Orthogonal P-wave morphology was accessed according to P-wave positive/negative deflection, or biphasicity, in leads X, Y, and Z. Three predefined types (orthogonal type 1, 2, or 3) indicative of the interatrial conduction route were considered [
15]. In types 1 and 2, leads X and Y are positive, and lead Z is either negative or biphasic, while in type 3 lead X is positive and lead Y is biphasic. (
Appendix A.1.8).
In conclusion, a total of fourteen SPWIs was taken into consideration; the eight conventional 12 lead ECG PWIs, along with the MVP score, a-IAB, p-IAB, and orthogonal types 1–3.
2.2. P-Wave Beat-to-Beat Analysis
2.2.1. Beat-to-Beat Classification into Main and Secondary Morphology
Orthogonal vectorcardiograms were further studied to accomplish B2B analysis. Signal processing was performed using MATLAB R2015a, The MathWorks, Inc., Natick, MA, USA. Following an automated signal pre-processing procedure, consisting of denoising and QRS complex detection, artifacts and ectopic beats were removed in a semi-automated manner. According to a methodology previously described [
9], where the existence of main and secondary P-wave morphologies was proposed, a clustering technique was used to classify P-waves into distinct groups of main, secondary, or other less frequent morphologies. Specifically, a signal window of 250 ms preceding every QRS complex, named P segment (P
SEG), was defined. Following a k-means clustering method, the optimal number of clusters, which better classify the P
SEGs was calculated, and every P
SEG was allocated to a cluster. The mean morphology of the cluster containing most P
SEGs was used as a template for the detection of the P-waves matching the main morphology. For the remaining P
SEGs, following the same process once again, an additional P-wave template was extracted, to detect the P-waves allocated to the secondary morphology. The beats that did not match any of the morphologies were considered as other morphologies. Therefore, the secondary morphology was not extracted merely as the mean morphology of the P-waves not matching the main morphology, but rather as the mean values of a group of beats highly correlated and truly forming a cluster. Finally, the percentage of P-waves matching the main and the secondary morphology in each lead was calculated, as well as the percentage of P-waves following simultaneously the main morphology in all three leads (
Appendix A.2.1.).
2.2.2. Time-Domain Analysis
Time features regarding P-wave position related to the QRS complex were calculated. Thus, the distance between P-wave onset, offset, and maximum value, and Q- or R-points were assessed for every beat allocated to the main morphology.
2.2.3. Time-Frequency Domain Analysis
The P-waves of the dominant morphology were further analyzed in a B2B aspect using continuous wavelet transform, an approach already been used for processing non-stationary signals such as ECG. The base wavelet used was the complex Morlet wavelet. According to a previous study of ours, where optimal frequency zone range was investigated, analysis was performed in three non-overlapping bands (low—L: 30–70 Hz, medium—M: 70–160 Hz, and high—H: 160–200 Hz) [
9]. Frequency features such as mean, median, and maximum energy, as well as time-frequency features, such as maximum energy location regarding P-wave onset, offset, or peak, were calculated (
Appendix A.2.2.).
Eventually, P-wave B2B analysis, consisting of classification into primary and secondary morphology, and time-domain and time-frequency domain analyses, in three different axes, in three individual frequency zones, and both mean value and coefficient of variation calculation of selected variables ended up to a total of 262 features (
Table 2).
2.3. Statistical Analysis
Statistical analysis data were expressed as mean ± standard deviation for continuous variables. The Mann–Whitney U test or the Fisher’s exact test was used, as appropriate, for statistical testing. Univariate logistic regression analysis was performed on statistically significant features to determine variables with prognostic value on detecting patients with a history of AF, and variables were ranked according to their area under the curve (AUC). Predictors were checked for collinearity, and the features with the highest AUC among those highly correlated were selected for further study.
Following multivariate logistic regression analysis, two different classifiers were developed—one based on features derived from B2B analysis and one using SPWIs. Receiver operating characteristics (ROC) curves were delineated for these classifiers, and the AUC was calculated to estimate the predictive value of each one. AUC comparison was accomplished using DeLong’s method [
16]. Statistical analysis was performed using MATLAB (R2015a) computer software, and an alpha level of <0.05 was accepted as statistically significant.
3. Results
The studied dataset consisted of 276 features in total; of those, 262 variables were obtained from B2B analysis and 14 from SPWIs. Five SPWIs were found significantly different between the two groups, while Β2Β analysis was represented by nineteen statistically significant features; seven of them were derived from the percentage of main or secondary morphologies and twelve from spectral and temporal analysis (
Table 3).
Nine traditional PWIs did not differ significantly between the two groups: PR duration, P-wave dispersion, P-wave peak time, P-wave axis, PTFV1, P-wave orthogonal type, and a-IAB values were comparable in control and PAF groups. On the other hand, the widely used P-wave duration was longer, and the P-wave area, along with P-wave voltage in lead I were smaller in PAF patients, compared to the control group. Moreover, p-IAB was more common and the MVP score was higher in the PAF group.
Additionally, morphology analysis revealed significant differences between main, secondary, and secondary to the main ratio in both axis X and Y, along with the percentage of P-waves following main morphology in all leads at the same time. Furthermore, both time-domain features and frequency-derived parameters in all three axes were represented by significant variables such as coefficient of variation of P-wave peak time in axis X, mean value of maximum energy in the high-frequency band in axis Y, or mean value of maximum energy location to P-wave duration in the mid-frequency band in axis Z.
Following univariate logistic regression analysis, variables were ranked according to estimated AUC, and odds ratios (OR) were calculated. Sixteen features were significant predictors, while eight variables did not reach the level of significance or AUC lower limit above 0.5 and were excluded from further analysis (
Table 4).
Collinearity test revealed a strong correlation between main, secondary, and secondary to main morphology ratio in axis X, between the coefficient of variation of the distance between P-wave onset and the Q- or R-wave in axis X, and between the MVP score and P-wave duration. Only the features with the highest AUC among those highly correlated were selected (
Table 4). Thus, eight B2B features and four SPWIs were finally used in multivariable regression analysis. Eventually, two classifier groups were developed: one using SPWIs and one using B2B variables. The risk of overfitting was kept low by ensuring an event-per-variable ratio above 10, and thus, only models consisting of ≤ 3 compounds were taken into consideration [
17].
All suggested models for each category were compared to each other in terms of AUC, and those with the highest performance in our data sample were selected (
Table 5). In the first case, a B2B classifier consisting of the following three features was proposed: (i) coefficient of variation of distance between the P-wave peak and Q-wave in Χ axis, (ii) percentage of P-waves allocated in main morphology in X axis, and (iii) mean value of maximum energy in the high-frequency band in the Y axis. In the second case, the model with the highest AUC among all SPWIs classifiers consisted of (i) p-IAB, (ii) P-wave voltage in lead I, and (iii) P-wave duration.
These two classifiers were finally compared to each other, and the B2B model was found superior to the SPWIs model: B2B AUC: 0.849 (0.754–0.917) vs. SPWIs AUC: 0.721 (0.613–0.813),
p value: 0.041 (
Figure 1).
4. Discussion
In the current study, we assessed the performance of B2B P-wave analysis in distinguishing PAF patients from individuals with no history of AF, compared to that of standard PWIs. In our sample, a B2B-derived model outperformed a classifier based on established ECG-based PAF predictors.
Univariate analysis revealed that only B2B indices, and none SPWI, achieved an acceptable AUC > 0.7 in predicting PAF (
Table 4). Furthermore, the B2B classifier consists of three components coming from three different fields of B2B analysis: main and secondary morphology analysis (percentage of P-waves allocated in main morphology in X axis), temporal analysis (coefficient of variation of distance between the P-wave peak and Q-wave in Χ axis), and wavelet analysis of main P-wave morphology (mean value of maximum energy in the high-frequency band in the Y axis), underlying the importance of an integrated approach to PAF prediction. Additionally, the SPWI classifier consists of p-IAB, voltage in lead I, and P-wave duration, features that, interestingly, are also compounds of the recently developed MVP score [
14].
Many PWIs have been previously found to be highly correlated to AF development. P-wave duration may be the most widely studied among them. In our study, P-wave duration along with voltage in lead I, and patrial IAB was one of the variables included in the SPWI classifier. In a meta-analysis of Framingham Heart Study (FHS) and Atherosclerosis Risk in Communities (ARIC) study, P-wave duration > 120 ms was significantly associated with AF [
18]. Moreover, low P-wave amplitude in lead I has been related to displaced interatrial conduction and AF recurrence in another study [
19].
Additionally, in the aforementioned work, the P-wave area did not show a robust correlation with AF, and PTFV
1 predicted AF only in the ARIC study and not in FHS [
18]. However, PTFV
1 has been found in other studies to be significantly associated with AF occurrence [
3,
12] but not in ours.
Advanced interatrial block (a-IAB) is an established electrocardiographic phenotype, that has been thoroughly studied. However, in Regicor Study it was found that a-IAB did not seem to provide an additional AF risk beyond that of P-wave duration [
20], which was concordant with our findings. Risk factors for developing a-IAB such as aging, hypertension, and obesity are similar to those for developing AF, interpreting the relationship between a-IAB and AF [
20]. Since, patients selected in our study were relatively young, with a short history of PAF and few other comorbidities, there were very few cases of a-IAB, not enough to support a statistically significant result. On the contrary, p-IAB was found to be a quite powerful predictor and thus was included in the proposed SPWI classifier.
One of the novelties in our work is the study of P-wave variability with means of B2B analysis. Most previous studies on P-wave analysis were conducted on 10 s ECG recordings or signal-averaged ECGs. However, a B2B approach on a P-wave study is feasible and effective [
21,
22]. In fact, in our study, not only we managed to observe P-wave variability in 10 min recordings, but we also succeeded in depicting and quantifying primary and secondary morphologies using an algorithm formerly described and repeatedly tested for consistency in various datasets. A higher percentage of recorded beats belonging to secondary morphologies and a lower percentage of beats of dominant morphology, in various orthogonal axes, were strongly associated with PAF, both in the current dataset and in a different dataset, previously studied [
9]. This P-wave variability can only be reliably quantified in relatively lengthy ECG recordings such as the ones we use in the current study and could be proved to be the main advantage of the proposed modality. In fact, since the whole process is largely automated, the workload required to analyze a 10-minute recording is considerably low.
In addition, it is of great importance to underline that the patients who participated in our work had a noticeably short history of PAF (less than a month), were relatively young (mean age: 55.4 ± 12.6), and had a low CHA
2DS
2-VASc score (1.2 ± 1.3). Thus, using the proposed analysis it is feasible to detect predisposition to AF in the early stages. Similarly, impaired interatrial conduction, assessed utilizing orthogonal morphology type, has been observed in young patients without comorbidities, implying that alterations in atrial electrophysiology may be the cause rather than the consequence of AF [
23]. Moreover, in our study, patients receiving antiarrhythmic drugs were excluded, so it is safe to conclude that there was no interference to results caused by antiarrhythmic medication.
The case–control design of this study is a relative limitation but considered acceptable for a study investigating a novel diagnostic modality. Larger sample size would probably provide more robust results. In this work, simple logistic regression was employed, but in the future, more sophisticated artificial intelligence-based classifiers trained with features derived from larger populations are expected to reach a higher level of accuracy.
Although a well-known ECG-derived AF predictor, signal-averaged ECG (SAECG) was not analyzed in the current study. Recording atrial late potentials is technically challenging; thus, the hypothesis that low-amplitude atrial potentials are associated with AF was initially rejected [
24]. However, using improved techniques, several SAECG P-wave indices were studied. In fact, atrial SAECG is widely used to calculate filtered P-wave duration (FPD), a more precise and reproducible measure of P-wave duration. On the contrary, SAECG late potential parameters, such as root mean square voltage of terminal P-wave segments, supposed to represent the activation of the left pulmonary veins [
25], have shown little reproducibility and their clinical importance is controversial [
26]. Therefore, in patients with PAF, FPD was longer, compared to normal controls, while no statistical difference was found in atrial late potentials [
27]. Furthermore, FDP can identify hypertensive PAF patients while in sinus rhythm [
28] and, along with left atrial dimension, is considered significant predictor of transmission from paroxysmal to persistent or permanent AF [
29,
30].
Platonov et al. have shown that fibrosis and fatty infiltration were more manifest in autopsies of patients with permanent AF than in those with paroxysmal AF, regardless of patients’ age [
31]. Moreover, P-wave duration and amplitude show a significant correlation with low-voltage area size and may be used as a non-invasive tool to predict severe scarring, as experimental and simulation studies suggest [
32,
33]. Therefore, PWIs probably predict AF progression by reflecting atrial fibrosis extension. Furthermore, B2B P-wave variability in patients with paroxysmal AF could be explained by variability in sinoatrial node exit location in combination with slow conducting regions, such as scars, as has been shown in computational simulation studies [
34,
35].
In future studies, an association between B2B indices and AF predictors other than SPWIs can be examined. Thus, a correlation between B2B analysis and echocardiographic parameters of left atrial size and function or characteristics of the left atrial substrate such as low voltage areas would be a potential research topic. Moreover, prospective studies can be conducted to determine B2B predictors’ ability to detect clinical conditions beyond AF, such as embolic strokes of an undetermined source (ESUS).