Learning optimal early decision treatment rules with multi-domain intermediate outcomes

Implementing precision medicine for mental disorders presents challenges due to disease complexity and heterogeneity in patient responses. Empirical studies suggest that early indicators, such as interim measures (e.g., interim patient self-reports) of disease improvement or relapse, can predict longer-term outcomes, serving as proxies when final outcomes (e.g., in-clinic assessments) are less accessible. However, existing approaches for deriving individualized treatment rules (ITRs) often ignore these early signals, instead focusing only on a final outcome as the reward. In this work, we propose a new method incorporating intermediate outcomes from various domains into a personalized composite outcome, serving as the reward for learning ITRs. This composite is a weighted sum of inferred latent states from observed measures, with weights personalized for each patient, ensuring consistency with the long-term final response. Our simulations show that this approach not only provides early detection of non-responders but also improves long-term treatment outcomes. Applying our framework to a randomized clinical trial on major depressive disorder (MDD) demonstrates its effectiveness and advantages in ITR learning.