A single-center prospective exploratory observational study was performed. Patients at the University Heart Center Dresden admitted for catheter ablation of symptomatic atrial fibrillation underwent a routine pre-procedural CMR for ablation planning. Consecutive patient data was collected from April 2019 until August 2020. The study was approved by the local ethics committee (EK 284092012). Informed patient consent was obtained.

Inclusion criteria for this study were symptomatic paroxysmal or persistent AF, first AF ablation procedure, assessment of left atrial bipolar voltage mapping during AF ablation in sinus rhythm, and pre-procedural CMR imaging. All patients admitted for primary AF ablation at our center were eligible for a pre-procedural CMR to evaluate the pulmonary vein anatomy. However, due to limited CMR capabilities, not all patients received a CMR prior to the procedure or did not receive additional adipose tissue image acquisition, which was an inclusion criterion for this study. No clinical criteria were applied to select patients for CMR to avoid a selection bias. Re-do procedures did not routinely receive CMR imaging since either a previous CT angiography was present or an electro-anatomical map was available from the previous ablation.

Patients were excluded from this study if permanent or valvular AF, any previous AF ablation, left ventricular ejection fraction <30%, corrected mitral valve disease, restrictive or hypertrophic cardiomyopathy, or left atrial appendage thrombus was present. Additionally, patients were excluded with missing or non-Dixon epicardial adipose tissue imaging or inadequate image quality. A study protocol overview is provided in Fig. 1.

Study design flowchart. Abbreviations: atrial fibrillation (AF), cardiac magnetic resonance (CMR), left atrium (LA), left atrial appendage (LAA), low-voltage zones (LVZ), left ventricular ejection fraction (LVEF), Dixon imaging as first described fat–water separation imaging by W. T. Dixon in 1984

The CMR protocol (Siemens Healthineers MAGNETOM Aera 1.5T) consisted of a gadolinium-based 3D contrast-enhanced MR angiography with real-time bolus tracking to evaluate left atrial and pulmonary vein anatomy, using a FLASH3D gradient echo sequence (single breath-hold, real-time bolus tracking, without ECG gating) for image acquisition with the following parameters: FOV 400 mm, matrix 269 pixels × 384 pixels, voxel size 0.5 mm × 0.5 mm × 1 mm.

Adipose tissue adjacent to the left atrium was visualized using a fat-water separation sequence based on the Dixon imaging method, resulting in fat-only images (Fig. 2) [41, 42]. Imaging parameters used were slice thickness 5–7 mm, spacing between slices 5–8 mm, FOV 420 mm, matrix 144 pixels × 256 pixels, voxel size 0.8 mm × 0.8 mm × 5–7 mm, TR 586 ms, TE 1.5 ms. The number of acquired short-axis slices ranged from 6 to 10 to cover the whole left atrium of varying dimensions.

Example of a short-axis fat-only Dixon image of epicardial adipose tissue on the level of the left atrium, dorsal to the mitral valve plane

Generation of 3D models and volume quantification were performed using ADAS 3D software [43] (Galgo Medical S. L, 2019). The functionalities include 3D left atrial semiautomatic segmentation and automatic classification of tissues using selectable imaging thresholds. An example segmentation workflow is shown in the supplement.

Presence and extent of LVZ were assessed during the pulmonary vein isolation procedure by bipolar voltage mapping of the left atrial wall using a multipolar mapping catheter of variable diameter size (15–25 mm) and 1 mm electrode/8 mm spacing (Lasso 2515 NAV Eco, Biosense Webster, Inc., Diamond Bar, CA, USA, or Advisor™HD-Grid, Abbott, Abbott Park, IL, USA). All point measurements were taken during sinus rhythm. Patients that presented with AF received an electrical cardioversion during the procedure. In case of unsuccessful initial cardioversion, isolation of the pulmonary vein ostia was performed and the cardioversion attempt repeated. Classification of a prespecified left atrial wall region as being a low-voltage zone (LVZ) was performed if a median voltage ≤0.5 mV was recorded [44].

Statistical analyses were conducted using IBM SPSS (version 26; IBM Corp., Armonk, NY, USA) and GraphPad Prism (version 9.4.0; GraphPad Software, San Diego, CA, USA). Continuous variables were tested for normality with the D’Agostino–Pearson test. Normally distributed data are reported as mean ± standard deviation (SD), while non-normal data are presented as median (interquartile range). Between-group comparisons of continuous variables employed the Student’s t-test or Mann–Whitney U test, as appropriate. Categorical variables are expressed as counts and percentages and compared using Fisher’s exact test.

Correlations between two normally distributed continuous variables were assessed by Pearson’s correlation coefficient; point-biserial correlation was used when one variable was dichotomous. Associations between two binary variables (e.g., gender and LVZ presence) were estimated using the tetrachoric correlation coefficient under a bivariate normal model. Robust 95% confidence intervals and two-tailed p-values were derived from 1000 bootstrap resamples.

Intra-observer and inter-observer reliability for imaging measurements was evaluated using intraclass correlation coefficients (ICCs). Intra-rater reliability for CMR-derived left atrial volume (LAV) and left atrial epicardial adipose tissue volume (LA-EATV) was assessed by having the same investigator remeasure absolute values in ten randomly selected patients, blinded to prior results and clinical data, at two time points approximately 3 months apart. Inter-observer reliability for LA-EATV was assessed by a second investigator in 12 randomly selected patients, blinded to previous results. The ICC (95% CI) was calculated from single measurements using a two-way mixed-effects model, absolute agreement definition. For inter-observer reliability, a secondary analysis of consistency was performed. ICC values were interpreted according to conventionally accepted thresholds [45].

Prior to regression modeling, multicollinearity among predictor variables was examined by calculating variance inflation factors (VIFs), with VIF >5 indicating potential collinearity concerns.

Predictive models were developed via binomial regression including main effects only, and their discriminative performance was assessed by receiver-operating characteristic (ROC) analysis. Regression coefficient stability was validated by bootstrap resampling (n =2000) to derive bias-corrected and accelerated confidence intervals.

Model improvement was quantified using the categorical net reclassification index (NRI) with predefined risk categories (< 10%, 10–20%, > 20%). The NRI reflects the proportion of events correctly reclassified upward and non-events correctly reclassified downward by the new model versus the reference. Ninety-five percent confidence intervals and p-values for the NRI were obtained via 1000 bootstrap replicates.

Unless otherwise specified, statistical significance was defined as p <0.05. In the binomial regression, predictors with p ≤0.20 were retained for model building.