Bayesian Machine Learning for Estimating Optimal Dynamic Treatment Regimes With Ordinal Outcomes

Dynamic treatment regimes (DTRs) are sequences of decision rules designed to tailor treatments based on patients' treatment history and evolving disease status. Ordinal outcomes frequently serve as primary endpoints in clinical trials and observational studies. However, constructing optimal DTRs for ordinal outcomes has been underexplored. This article introduces a Bayesian machine learning (BML) framework to estimate optimal DTR with ordinal outcomes. To deal with potential nonlinear associations between outcomes and predictors, we first introduce ordinal Bayesian additive regression trees (OBART), a Bayesian tree-based model that integrates the latent continuous variable framework within the traditional Bayesian additive regression trees (BART). We then incorporate OBART into the BML (named BML-OBART) to estimate optimal DTRs based on ordinal data and quantify the associated uncertainties of the estimated parameters. Extensive simulation studies were conducted to evaluate the performance of the proposed method in comparison with existing methods. We demonstrate the application of the proposed BML method using data from a smoking cessation trial and provide the OBART R package along with R code for the implementation of BML-OBART.