Fig. 3

Receiver operating characteristic (ROC) analyses for unfavorable hemorrhage predictive models using all variants in regression model (blue curve), the bAVM classification system (green curve), and concise regression system including significant risk factors from the full model (red curve) (reference line shown in teal). The predictive accuracy of bAVM classification system (ROC areas of 0.77) was not less than that of full regression and concise regression model (ROC areas of 0.76 and 0.72, respectively; pairwise comparison of ROC curves, P = 0.87 for bAVM classification system versus full regression model and P = 0.08 for bAVM classification system versus concise regression model)