The study was initiated after receiving approval from the Rambam Health Care Campus Institutional Review Board and Ethics Committee on Human Research (Approval ID: RMB-0057-15). The need for written informed consent was specifically waived by the because of the retrospective nature of the study.

From our database of patients who had undergone cardiac CT between January 2012 and December 2019, we retrospectively identified consecutive patients who had undergone a cardiac CT examination using spiral scanning with retrospective gating, Scanning was performed on a dual-source Somatom Definition Flash scanner (Siemens Healthcare), with a temporal resolution of approximately 70 ms. Data was reconstructed every 5% of the cardiac cycle to provide 20 phases per study. We excluded patients with atrial fibrillation during the study, unstable patients, and insufficient CT quality.

Data were analyzed on a dedicated CT workstation (IntelliSpace Portal, version 11, Philips Healthcare) using the CT Comprehensive Cardiac Analysis software (Philips Healthcare). Retrospective data were analyzed using fully automatic segmentation of the heart chambers, producing phasic volume curves of each heart chamber. LA volume-based indices of LA function were calculated as previously described [15]. The agreement between the automated and manual measurements was high (Supplementary Material).

The primary endpoint of the study was incident AF after the cardiac CT. Incident AF ascertainment was accomplished by a comprehensive review of the medical records of all study participants.

The study participants were divided into three groups: Patients without AF at any time, patients with a history of AF prior to the cardiac CT, and patients without a history of AF but with incident AF after the cardiac CT. We first analyzed the association between measures of LA volume and function with previous history of AF. Then, after excluding patients with a history of AF, we studied the association between measures of LA volume and function with incident AF.

Maximum LA volume was defined as LA volume at end-systole just before mitral valve opening and was indexed to the body surface area to derive the LA volume index (LAVI).

Indexes of global LA function, LA reservoir function, early passive filling and LA booster function are summarized in Supplementary Table 1. Global LA function was assessed by LA total emptying function (LATEF) and LA function index (LAFI), which normalizes function to stroke volume and is rhythm-independent [8, 12, 13, 16,17,18,19]. Measures of reservoir function included reservoir volume index (ResVi) and LA expansion index (LAEI), representing the relative LA volume changes during the reservoir phase.

Early passive LA function was characterized by passive LA emptying fraction [15, 20, 21]. LA booster function was characterized by LAEFBooster [20, 22].; Other surrogates of LA booster function included LA booster contribution to LV stroke volume.

Based on previous studies, LA enlargement was defined as LAVI ≥ 62 mL/m2 [14], and normal global LA function was defined as LATEF ≥ 40% [13, 19].

Continuous variables are presented as mean ± SD or medians (with interquartile ranges), and categorical variables as numbers and percentages. The relationship between various parameters of LA volume and function were assessed by Pearson correlation.

The associations between LA function and clinical characteristics and CT data were assessed with the use of univariable linear regression. Variables found to have a univariable association with LA function at the p ≤ 0.1 level (Wald test) were used in multiple linear regression. Partial correlation coefficients were used to estimate the proportion of variance of dependent variables explained by each independent variable.

Kaplan–Meier plots were used to assess the event-free probability of AF associated with LA volumes and function, and the log-rank test was used to compare the curves. The association between LATEF and the endpoint of incident AF was then analyzed using Cox regression analysis. The following clinical covariables known to predict AF [23, 24] were used in a multivariate stepwise Cox proportional hazards model: age, sex, body mass index, previous heart failure, hypertension, diabetes mellitus, coronary artery disease, multivessel coronary artery disease, previous myocardial infarction, previous coronary artery bypass surgery, left ventricular ejection fraction, and severe aortic or mitral valve disease. In addition, measures of LA size and LA function, dichotomized above or below the upper normal value or used as continuous variables, were considered in the model. To address the risk of model overfitting, we also computed a more parsimonious multivariate Cox model that included fewer variables (Wald test; p < 0.05 in the univariate analysis).

The incremental value of LA volume and function over common clinical predictors of AF was assessed by exploring changes in the global chi-square values in sequentially constructed multivariate Cox models. We compared models using the Akaike information criterion, Bayes information criterion (likelihood measures in which lower values indicate better fit and in which a penalty is paid for increasing the number of variables in the model), and Harrell C index. Comparison of C indexes was made as proposed by Newson [25].

Restricted cubic spline transformations of the continuous independent variable were implemented to detect a potential nonlinear relationship of continuous variables, using three knots placed at default locations. Statistical significance of nonlinearity (i.e., curvature) was tested by comparing the cubic spline model with the linear model, and p-values of < 0.05 were regarded as statistically significant nonlinear relationship between the exposure and the outcome.

To determine whether LA dysfunction mediates part of the association between LA size and risk of AF, we performed a causal mediation analysis where LA volume was the independent covariable, LA function was the mediator, and incident AF was the dependent variable. Attenuation of the regression coefficient for LA volume in the mediation model compared with the logistic regression model provides qualitative evidence for mediation [26]. We used a 2-way decomposition and estimated the multivariable-adjusted direct and indirect effects (i.e., mediation effect) and percentage (%) mediated by LA function. Multivariable adjusted (for all predictors of AF) direct and indirect effects (i.e., mediation effect) are reported, with calculation of 95% CIs using bootstrapping with 500 resamples. Statistically significant mediation was determined if the indirect effect was significantly different from zero.

Differences were considered statistically significant at the 2-sided p < 0.05 level. Statistical analyses were performed using STATA Version 18.0.