Figure 1.

(A, B) legend: The receiver-operator characteristic (ROC) curves for the prognostic model. The ROC curves had an area under the curve of 0.85 in-sample and 0.91 out-of-sample, indicating good discrimination between good and poor outcomes. (C, D) legend: The calibration curves for the prognostic model. Predicted probabilities of poor outcomes were split into five risk quintiles and plotted against observed probability of poor outcomes in that quintile to determine model calibration to observed risks. The points that fall along the dashed 45-degree line indicate perfect agreement between the predicted probability and the observed probability of poor outcomes in the data. Error bars are ± 1 standard deviation of draws from the posterior distribution of the outcome. While agreement between the predicted and observed data appears strong for lower-risk quintiles, there were some deviations as the observed probability of poor outcomes increased.