Prediction of the Best Treatment Assignment Using Random Forest with Regression in the Nodes

The goal of personalized medicine is to identify treatment with the highest expected success given individual factors. For many diseases, such as substance use, a reductionist approach, i.e. finding a small set of strong predictors, might not be feasible due to a large number of factors influencing the outcome. Rather than aiming for a single “best” model we consider a collection of models allowing contribution of weak predictors. Although random forests provide stable predictions, they only produce simple data summaries (e.g. prevalence) in the nodes and thus don’t allow consideration of “what if” scenarios. Recently we have developed methodology (mobForest available through CRAN contributed repository) that combines random forest methodology with regressions in the nodes. Thus, for an individual with a specific combination of factors our approach provides smooth and stable prediction of potential treatment outcomes under several alternative treatments. The models are validated using independent training/test datasets. We illustrate the use of the tool on a dataset from the largest clinical trial of several alcohol treatment approaches called COMBINE. We show that while one of the treatments is best for most of the subjects there are individuals who could benefit from alternative treatments.