Abstract:
|
Individualized treatment rules (ITRs) are gradually being considered to replace "one-size-fits-all" strategy to personalize medical decision making. In this paper, we propose a machine learning approach to estimate ITR applicable to both observational studies and randomized clinical trial (RCT), referred as matched learning (M-Learning), which proposes to perform matching instead of inverse probability weighting to alleviate confounding in observational studies. In addition, a matching function is proposed to compare outcomes for matched pairs where discrete outcomes can easily be handled. We further improve efficiency of estimating ITR by using stratification based on predictive scores to construct informative matched pairs. The advantage of M-learning includes improved robustness, stability, accuracy, and flexibility to accommodate complex patterns among features. We prove Fisher consistency of M-learning, conduct extensive simulation studies and show M-Learning outperforms existing methods when propensity scores are misspecified and in certain scenarios of presence of unmeasured confounders. Finally, we apply our method to an RCT and electronic health records (EHR) data.
|