Construction and evaluation of AD predictive models in different independent datasets (ADNI, ROSMAP, GSE5281). Random forest algorithm was applied on the three different cohorts to establish AD predictive models through 10-fold CV (ADNI (A), ROSMAP (B), GSE5281 (C)). DeLong’s tests were used to compare the difference between ROC curves. Kaplan–Meier survival curves were used to estimate the accuracy of the hub genes signature on predicting survival. Individuals were well stratified by their risk predictions, and individuals in the high-risk group had a significantly earlier age of AD onset, as found for all ADNI (D), ROSMAP (E), GSE5281 (F) datasets (Log-rank test, p < 0.0001).