An individualized treatment rule (ITR) represents a decision function that assigns an optimal treatment based on each individual’s covariates. Especially in ultra-high- dimensional ITR, removing irrelevant variables is important both in terms of cost and interpretability. To this end, several variable selection methodologies have been proposed, notably marginal screening techniques, which are one of the most model-free methods. However, these methods suffer from the problem that they cannot cope with the special structure of the data. On the other hand, Random Projection Ensembles, a type of dimensionality reduction method, have demonstrated excellent performance in various fields. Therefore, in this study, we propose a new algorithm for feature screening in ITR, based on the concept of Random Projection Ensembles. This method is designed to be more adaptable to the special structures of data, potentially overcoming the issues associated with traditional marginal screening techniques. Our paper will detail the development of this algorithm and demonstrate its application in making ITR more efficient and easier to interpret, especially in cases with a large number of variables.