Fig. 5

For each PBCG cohort as an individual test set, all other 9 PBCG cohorts were used as a training set to fit a model, which was subsequently evaluated by the AUC, HLS and net benefit at the 15% threshold. The process was then repeated for each test set using the other 8 PBCG cohorts excluding Zurich as a training set. The AUC difference is reported along with 95% confidence intervals. For the negative of HLS and net benefit at 15% threshold, median estimates of difference and 95% percentile intervals (2.5 to 97.5 percentile) are obtained via bootstrapping. Positive values indicate inclusion of Zurich improves the respective performance characteristic for the test set