Fig. 1

Recursive RF feature selection process. The feature selection process started with a large dataset: all CpGs that survived data cleaning and preprocessing, and were not potentially affected by probe SNPs. The cycle in black (conducting the Random Forest, collecting evaluation measures, assessing stop criteria, and reducing the data) repeated until the atopy-specific misclassification rate showed a marked increase, indicating that some excluded sites were important in classifying atopic participants. Thus, once an increase in atopy-specific misclassification was observed, the cycle stopped and sites from the previous iteration were selected for follow-up testing. OOB-ER out-of-bag error rate, RF Random Forest, VIM variable importance measure