SET-VALUED DYNAMIC TREATMENT REGIMES

Purpose: The goal of personalized medicine is to use patient characteristics to recommend a treatment that leads to a favorable clinical outcome. A common attempt at realizing this goal is to use data to construct a decision rule that takes patient characteristics as input and provides a single recommended treatment as output. This approach is misleading in cases where there is not enough evidence to discriminate between two or more effective treatments.

Method: We propose a method for constructing a decision rule that provides a class of recommended treatments. The new, potentially multivalued decision rule allows clinicians to select from among the recommended class of treatments based on considerations such as individual preference, clinical expertise, cost, and local availability. Classes of recommended treatments are formed by effectively “fusing together” like treatments using a lasso-type penalty. This approach is shown to be particularly appealing when applied to multistage decision making problems where the sequential nature of the decision rule complicates traditional testing type approaches.

Result: Theoretical and empirical evidence suggest that multivalued decision rules can lead to better expected clinical outcomes than traditional single valued decision rules when these rules must be estimated from noisy data. Moreover, using the Study of the Adaptive Interventions for Children with ADHD Trial we illustrate how to construct and interpret multivalued decision rules.

Conclusion: Multivalued decision rules balance empirical evidence and clinical expertise by screening out treatments that are suggested by the data to be inferior and leaving the clinician to choose from among the remaining treatments.