Prevalence and determinants of atrial fibrillation progression in paroxysmal atrial fibrillation

Objective Atrial fibrillation (AF) often progresses from paroxysmal AF (PAF) to more permanent forms. To improve personalised medicine, we aim to develop a new AF progression risk prediction model in patients with PAF. Methods In this interim-analysis of the Reappraisal of AF: Interaction Between HyperCoagulability, Electrical Remodelling, and Vascular Destabilisation in the Progression of AF study, patients with PAF undergoing extensive phenotyping at baseline and continuous rhythm monitoring during follow-up of ≥1 year were analysed. AF progression was defined as (1) progression to persistent or permanent AF or (2) progression of PAF with >3% burden increase. Multivariable analysis was done to identify predictors of AF progression. Results Mean age was 65 (58–71) years, 179 (43%) were female. Follow-up was 2.2 (1.6–2.8) years, 51 of 417 patients (5.5%/year) showed AF progression. Multivariable analysis identified, PR interval, impaired left atrial function, mitral valve regurgitation and waist circumference to be associated with AF progression. Adding blood biomarkers improved the model (C-statistic from 0.709 to 0.830) and showed male sex, lower levels of factor XIIa:C1-esterase inhibitor and tissue factor pathway inhibitor, and higher levels of N-terminal pro-brain natriuretic peptide, proprotein convertase subtilisin/kexin type 9 and peptidoglycan recognition protein 1 were associated with AF progression. Conclusion In patients with PAF, AF progression occurred in 5.5%/year. Predictors for progression included markers for atrial remodelling, sex, mitral valve regurgitation, waist circumference and biomarkers associated with coagulation, inflammation, cardiomyocyte stretch and atherosclerosis. These prediction models may help to determine risk of AF progression and treatment targets, but validation is needed. Trial registration number NCT02726698.

arteries. The aortic PWV was determined by using ≥ 20 consecutive pressure waveforms at the carotid and femoral artery. The wave transit time was calculated by the system software using the R-wave from the simultaneous ECG recording. of the simultaneously recorded ECG. Distance between both measure points was determined and corrected by multiplying the distance by 0.8. The PWV was calculated by dividing the distance between the femoral and carotid artery by the wave transit time. IMT and presence of plaques was assessed by ultrasound (Siemens Acuson S2000) with the Syncho US Workplace 3.5, Arterial Health Package for automated IMT measurement. Assessment of the IMT was done bilaterally in the common carotid artery, the carotid bifurcation, and internal carotid artery.
Cardiac computed tomography (CT) Epicardial fat was measured on ECG-triggered, native CT heart scans according to the methodology introduced by Fox et al.
(1). Tube voltage of scan protocols varied between 80-120kV. The region of interest (ROI) was defined as described by Versteylen(2): The cranial slice limit was set at the level of the 3 BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s) carina of the pulmonary artery, and the caudal slice limit was the last slice containing any portion of the heart. The anterior border of the ROI was defined by the sternum, the posterior border by the ribs and vertebral column. Images were reconstructed using a soft-tissue algorithm. The pericardium was traced by a blinded reader placing 5-7 control points per slice using axial views as described earlier. Afterwards Catmull-Rom cubic spline functions are then automatically generated to obtain a smooth closed pericardial contour. Ultimately fat was automatically summed with a dedicated volumetric software (syngo.via Frontier, Cardiac risk assessment package, Siemens Healthineers, Forchheim, Germany).
Epicardial and pericardial fat were defined as previously described by Iacobellis: Epicardial fat is located between the outer wall of the pericardium and the visceral layer of the pericardium. Pericardial fat is localized between visceral and pericardial myocardium(3). 3. Iacobellis G. Epicardial and pericardial fat: close, but very different. Obesity (Silver Spring).

Blood biomarkers
At baseline peripheral blood samples were collected. Patients needed to be in sinus rhythm during blood sampling and oral anticoagulation was temporarily interrupted. All blood samples were processed and stored at -80°C.

4
BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s)

Extensive statistical description
Fisher's exact was used for binary variables and the T-test and Wilcoxon test were used depending on normality for continuous variables. For non-binary categorical variables, the Chi-squared test with simulation Multivariable logistic regression model Collected baseline variables including core lab data, with p<0.10 in the age and sex adjusted logistic regression, with exception of EHRA class, number of comorbidities, CHA 2 DS 2 -VASc score and medications, were included in a bidirectional step-wise variable selection, starting with age and sex in the model. Variables were then added to the model in order of (increasing) p-value of age and sex-adjusted analyses, starting with the variable that has the lowest p-value in the age and sex-adjusted logistic regressions. Before a new variable is added, variables in the model with p>=0.05 were identified and removed, starting with the one with the biggest p-value. Before each potential next removal, the model was refit and thus the recalculated p-value was used to determine if there was a next variable with p>=0.05. If no variables were to be removed the next variable was added. The bi-directional stepping consists of a single forward stepping of all the variables, interspersed with backwards stepping of the variables in the model before each next step in the forward-stepping. The statistical criterion for removing a variable from the model (during each backwards stepping) was p>=0.05. The step-wise variable selection will ensure that no variables with p>=0.05 will end up in the final multivariable model. However, due to possible negative confounding even a variable with p>=0.05 in an age and sex-adjusted model, may have p<0.05 in a model with additional covariates. The aim was also not to keep a too big set of variables to be included in the step-wise process. Therefore, a p<0.10 was selected as a trade-off between the most stringent selection (based on p<0.05) and the least-stringent selection (that is without taking into account the p-value of age and sex-adjusted regressions).
In the final multivariate model (obtained at the end of the bidirectional stepping), testing for each possible second-order interaction was done (i.e. an interaction between two variables or an interaction with itself (quadratic term), what happened if this interaction is added to the model). Specifically, the p-value of the interaction was checked. Of all possible interactions, none reached Bonferroni significance (taking into account multiple testing).
Hosmer and Lemeshow test was used for goodness of fit test with 8 degrees of freedom, Chi-squared = 6.68, p-value = 0.572 Discrimination slope of the main model = 0.082 Imputation Imputation was implemented for missing values using the R package mice. Mice creates multiple imputations for multivariate missing data. For this article 4000 imputations for each model fit that required imputation were performed. Each incomplete variable was imputed by a separate model. The default methods of predictive mean matching for numeric data and logistic regression for binary data. For each logistic regression "massive imputation" was performed, which means that all variables in a model were at the same time also used for the imputation needed for the fit of that model. Internally, mice performs the logistic regression fit on all 4000 imputations. It pools the results according to Rubin's rules for imputation, with a small sample refinement of the method to compute degrees of freedom according to Barnard and Rubin.

Internal validation
Internal validation was accessed using bootstrapping.(1) Fifty bootstrap samples were used.(2). The optimism caused by overfitting in the C-statistic of our model without biomarkers to be 3.03%." Risk score The multivariable model was used to calculate the linear predictor for all patients with complete data. This was done in the standard way, namely linear combinations of the variables with the beta coefficients as weights, specifically we obtained the following expression for the linear predictor (centering continuous variables on their respective means): Linear prediction.= Female sex * (-0.5586208) + (PR-interval -168.8658) * 0.01309933 + (LA contraction function -17.30064) * (-0.08169752) + (waist circumference -101.1294) * 0.02787357 + mitral valve regurgitation * 1.771005 Next, these linear predictors was used to calculate a factor (called F here for simplicity), such that the 95% interval of the linear predictors (from the 2.5% quantile to the 97.5% quantile), when multiplied with this factor, is of length 10. This was done, because the aim was a point-based risk score that can vary from 0 up to and including 10 with only very few occurrences outside of this interval. F can be interpreted as a conversion factor such that the product of a variable with its beta coefficient and F represents a certain number of points. This factor F was then used to obtain a preliminary scoring scheme in the following way: For binary variables, the no-level gets zero points and the yes-level gets beta * F points. The number of points was rounded to the nearest integer value. For continuous variables, the step size is defined as the inverse of the absolute value of its beta coefficient and F rounded up to the nearest integer value.
Step size can be interpreted as the number of units of the variable per point. The number of levels the variable has in the point-based risk score is then set to the range of the variable (maximum minus minimum) divided by the step size rounded down to the nearest integer value. The range of the variable is then divided using intervals of length equal to the step size and centered in the range of the variable. The number of points assigned to the interval is the value of the variable in its midpoint times its beta coefficient times F, rounded to the nearest integer value. Finally, the first interval is the extended to include also the smallest values of the variable and the last one to include also the largest values, so that the entire range of the variable is covered. Age was removed from the point based risk score. Progression is presented as percentage 100×(P/ (tstoptstart ))% tstart = start time of the period of which the weighted AF burden is calculated, tstop = end time of the period of which the weighted AF burden is calculated.

Mathematical formula of AF progression
An increase >3% AF burden over the first six months or total follow-up was chosen as definition for atrial fibrillation progression. This cut-off point was chosen because the results were most consistent with the assessment of the physicians. Figure S1. RACE V study design overview Use of rhythm control therapy. AAD= antiarrhythmic drug; AF= atrial fibrillation; ECV = electro cardioversion; PVI = pulmonary vein isolation 2 patients with AF progression used amiodarone, both started during follow-up, 1 stopped during follow-up, 1 continued until end of analysis. 4 patients without AF progression used amiodarone, all started during follow-up, 1 stopped during follow-up, 3 continued until end of analysis. 5 of 26 patients (19%) undergoing PVI showed AF progression. 9 of 30 patients (30%) undergoing ECV showed AF progression. 3 of 14 (21%) patients undergoing both ECV and PVI showed AF progression.

Supplementary
10 BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s) 2 (4%) 10 (3%) 2 (12%) 3 (9%) Data are presented as mean±standard deviation, number of patients (%), or median (interquartile range). Abbreviations: ACE=angiotensin-converting enzyme; AF=atrial fibrillation; BMI=body mass index; CCA= common carotid artery; eGFR=estimated glomerular filtration rate; EHRA= European Heart Rhythm Association class for symptoms; HFpEF= heart failure with preserved ejection fraction; HFrEF= heart failure with reduced ejection fraction; IMT=intima media thickness; NOAC= novel oral anticoagulation; NT-proBNP=N-terminal pro-brain natriuretic peptide; *Atherosclerosis is presence of history of myocardial infarction, percutaneous coronary intervention, coronary artery bypass graft, ischemic cerebral infarction, peripheral vascular disease, Agatston score >400 or plaque; **The number of comorbidities was calculated by awarding points for hypertension, heart failure, age >65 years, diabetes mellitus; coronary artery disease, BMI>25kg/m2, moderate or severe mitral valve regurgitation and kidney dysfunction (eGFR<60); ***The CHA2DS2-VASc score assesses thromboembolic risk. C=congestive heart failure/LV dysfunction, H=hypertension; A2=age ≥75 years; D=diabetes mellitus; S2=stroke/transient ischemic attack/systemic embolism; V=vascular disease; A=age 65-74 years; Sc=sex category (female sex). ). a Left atrial and ventricle strain measurements could not be performed in 75 patients. Measurements of right atrial strain could not be done in 123 patients. b Agatston score was not available for 10 patients, epicardial and pericardial fat could not be analysed for 21 patients. c IMT CCA was not available for 55 patients, IMT all segments for le for 56 patients and pulse wave velocity could not be measured in 78 patients and amount of plaques could not be measured in 145 patients.