摘要：Personalized medicine has recently received increasing attention because of the significant heterogeneity of patient responses to the same medication. The estimation of optimal individualized treatment regime or individualized treatment rule is an important part of personalized medicine. Individualized treatment regimes are designed to recommend treatment decisions to patients based on their individual characteristics and to maximize the overall clinical benefit to the patients. However, most of the existing statistical methods are mainly focus on the estimation of optimal individualized decision rules for the two categories of treatment options and rely heavily on data from randomized controlled trials. There has been a relative lack of research work on the selection of multicategorical treatment options in real-world settings. We address this problem and propose a machine learning approach (CM-learning) to estimate optimal treatment regimes. This new learning approach allows for more accurate assessment of individual treatment response and alleviation of confounding, more importantly, CM-leaning is doubly robust, efficient and easy to interpret. We first introduce the concordance-based value function that measures weighted concordance for each patient by matching imputation. We then find the optimal treatment regime to maximize the concordance-based value function through the use of tree structure that directly handles the problem of optimization with multicategorical treatment options. Furthermore, an extension of CM-learning can be applied to ordinal treatment settings. Through a large number of simulation studies, we demonstrate that CM-learning outperforms existing methods. Lastly, the proposed method is illustrated in an analysis of AIDS clinical trial data.