Fig 1.

Most important predictors of AF according to: (A) the CHARGE-AF risk model; (B) the fitted baseline neural network; and (C) the fitted time-varying neural network. Variable importance was determined on the training dataset according to the absolute size of the published regression coefficients for the CHARGE-AF risk model[11] and Garson’s algorithm[23] for the fitted neural networks. Importance is expressed as a percentage of the most important predictor within each model. Importance is shown for all 11 variables in the CHARGE-AF risk model and the top 20 most important variables in each of the fitted neural network models.