The Backward elimination method
To select features automatically we iteratively fitted random forests,
at each iteration building a new forest after discarding 20% of the
features with the smallest variable importance. The selected set of
features was used as a predictor to fit the model to check the ‘out of
bag’ (OOB) error rate. We examined the OOB error rates from all fitted
random forests. We chose the solution with the smallest number of
variables whose error rate was within one standard error. This procedure
was performed iteratively using the varSelRF package in R (v3.6.1).