Extended Data Fig. 2: Holdout validation of daGOAT, PeriHSCT-NB, PeriHSCT-RF, and XGBoost in the adult cohort.
From: Dynamic forecasting of severe acute graft-versus-host disease after transplantation
a, Temporal profiles of AUPRC from day 1 through day 100. The bottom panel shows how the number of remaining positive cases (severe aGVHD cases that had not had onset) decreased over time in the test set. (Blue, daGOAT; orange, PeriHSCT-NB; red, PeriHSCT-RF; green, XGBoost.) b, Precision-recall curves for the four models on day 23 (daGOAT’s peak performance day for the adults according to internal validation) in the test set. c, Receiver operating characteristic curves for the four models on day 23 in the test set. (The dotted line represents the null model.) d, Cumulative incidence curves for severe aGVHD within the ensuing two-week window (days 24–37) after the test-set patients were stratified on day 23. All-cause death was treated as a competing event that precluded severe aGVHD when calculating the cumulative incidence curves for severe aGVHD. ‘High-risk’ (purple), daGOAT score among top 1/6; ‘low-risk’ (cyan), daGOAT score among bottom 5/6.
