Fig. 5

Validation results for four independent glioma datasets. a The IDH mutation classifier’s performance increased with prediction confidence on four validation sets, confirming the utility of model calibration. b ROC curves show that our LR model performs well on four separate validation sets. c Besides AUC, our model performs well four additional metrics. Recall is significantly higher than precision across validation sets, indicating that the model performs better on IDH-mutant astrocytomas than IDH-wildtype tumors. L Misclassified IDH-wildtype tumors are significantly younger than their correctly classified counterparts in three validation datasets (Mann–Whitney U). e, f Misclassified IDH-wildtype tumors in the TCGA validation set and a dataset published by Jonsson et al. tend to have better outcomes than correctly classified IDH-wildtype diffuse gliomas. g, h Plots of the results of two Cox proportional hazard models of histological WHO grade 4 IDH-wildtype glioblastomas that incorporated our model’s IDH mutation prediction, EGFR amplification, and + 7/ − 10