Fig. 3

Microbiome-based random forest classifier for prediabetes. Performance of RF classifier (trained on OTU abundance) distinguishing between NG and PD samples from India and Denmark. [Note: The data was split into training and independent test sets (in the ratio of 66:34) and a feature selection step adopted while training the model performed with 10-fold cross-validation (× 10 replicates); Top 10 features were selected from each cross-validation fold and ranked based on their cumulative importance (gini score used). Final “bagged” RF model was built using a set of features providing best training AUC, selected through progressively adding the ranked features into the model (up to a maximum of 100), while evaluating training AUC]