Figure 3.

Performance on internal-external cross-validation for each held-out hospital. Each point is a hospital. Relative size of the points is proportional to the number of admissions at that hospital. The Brier Skill score was calculated using the Brier score from the observed mortality rate of the held-out hospital as the reference. E95 denotes the 95th percentile absolute vertical distance between the calibration curve and perfect calibration. Note that ICI and E95 are in the same units, but the y-axes are scaled differently.