Fig. 5. Unreliable predictions in the Stroke risk model.

a Calibration curves, b normalized Brier scores (inset shows expanded region corresponding to $0.95\le \mathrm{Brier}∕{\mathrm{Brier}}_{\mathrm{null}}\le 1.02$), and c AUCs for the most untrustworthy predictions (red) and those that fall within the bottom 99% $U\left(\stackrel{\rightharpoonup }{x}\right)$ values (black). **p « 0.0001. Error bars are from 100 bootstrap splits and show one standard deviation (a) or one standard error of the mean (b and c).