Fig. 2: Performance of the prePMF/ET classifier.

a, b Upon the training cohort of patients treated at University of Florence, the model achieved mean area under the receiver operator characteristic (AUROC) of 0.90 ± 0.04 and an average precision (AP) of 0.91 ± 0.05 across 5-fold cross validation. c, d Upon external validation of patients treated at Moffitt Cancer Center, the prePMF/ET classifier achieved an AUROC of 0.88 and AP of 0.81. Baseline AP of the external cohort is 0.23, representing the proportion of prePMF diagnoses in the imbalanced cohort.