Abstract Body (Do not enter title and authors here): Background: Atrial fibrillation (AF) is often asymptomatic and detection of AF is challenging. 2D-Echocardiogram (2DE) provides a comprehensive assessment of cardiac structure and function, and hence, may be used to predict AF.
Aim: To derive a 2DE-based risk model to predict AF using supervised machine learning (ML).
Methods: We included 5,445 older adults from the Atherosclerosis Risk in Communities (ARIC) study who underwent transthoracic 2DE at baseline (visit 5, 2011-2013) and had no known history of AF. Individuals were randomly allocated to the training and testing sets in a 7:3 ratio. Incident AF was ascertained from hospitalization records and death certificate during follow-up. Missing values of 2DE measures with <10% missingness were imputed 50 times via Multiple Imputation by Chained Equations (MICE), and those with ≥10% missingness were excluded from the analysis. The Lasso approach was used for variable selection based on the consensus of the 50 imputed datasets. Significant variables (P<0.05) were used to build a model for prediction of incident AF in the training set by Cox proportional hazards regression. Harrell’s C-index was used to assess model performance.
Results: The median (IQR) age was 74 (71-79) years (59% female, 23% Black). During a median (IQR) follow-up of 7.0 (5.4-7.8) years, 738 participants developed AF. The cumulative incidence of AF was 1.4%, 5.0%, and 9.6% at 1, 3, 5 years, respectively. The table shows the hazard ratios for AF of each predictor in the final model. The Harrell’s C-index at 1-, 3-, and 5-year was 0.80, 0.72, and 0.73, respectively in the training set and was 0.76, 0.74, and 0.75, respectively in the testing set.
Conclusion: Our ML derived 2DE-based risk model has good model discrimination for prediction of AF, setting the stage for a deep learning model using 2DE images for AF prediction.