Abstract:
|
Dynamic treatment regimes (DTRs) are sequential decision rules for individual patients that can adapt over time to an evolving illness. The goal is to find the DTRs tailored to individual characteristics that lead to the best long term outcome if implemented. To facilitate implementation, it is important that the DTRs are easy to interpret. If a variable is important, it should contribute to the decision rule at all stages, and the effects should be consistent over stages. We introduce a general learning framework on stabilized dynamic treatment regimes (SDTRs), where we can make stabilized decisions over time using the repeated measured information on the same variable. The proposed method is based on directly maximizing an estimator of the expected long-term outcome over all constrained DTRs. Experimental studies showed a superior performance of the proposed method.
|