Primary Submission Category: Heterogeneous Treatment Effects
Active Feature Acquisition in Precision Medicine
Authors: Michael Valancius, Michael Kosorok, Junier Oliva,
Presenting Author: Michael Valancius*
A fundamental goal in precision medicine is to learn individualized treatment rules (ITRs) that use rich biomarker data to tailor treatment recommendations. When treatment effects are heterogenous with respect to demographic, genetic, or clinical data, these personalized actions can lead to improved health outcomes. While this motivates learning ITRs that are functions of rich and complex data, the collection of covariate information in routine practice is often dynamic and bears a cost (financial, time, inconvenience, etc.). Therefore, paradigms basing ITRs on static sets of features can have limited real-world applicability. In this work, we consider tailoring the feature acquisition process to an individual. We provide conditions under which currently acquired features are informative for deciding whether (and which) additional covariates would be beneficial to collect. Furthermore, we show that the objective function representing the expected health outcomes under a given a feature acquisition policy and an ITR can be cast as a weighted classification problem, enabling the usage of standard machine learning methods to learn these two policies. Simulation results demonstrate the empirical improvement of this approach compared to alternative approaches that do not tailor the feature acquisition process based on the observed individual characteristics.