In diabetes treatments, it is difficult for the medical specialists to decide an optimum treatment selection even though they refer to “Practice Guideline for the Treatment for Diabetes”. This study aims to assist diabetes treatments using machine learning. First, we create models using machine learning in order to classify whether a patient keep HbA1c below certain level or not. Next, we build a HbA1c simulator using these models. We train and evaluate the models using data from 1,994 patients. As a result, we able to develop the simulator that show how prescription changes patient HbA1c in next 120 days less than 2 errors with 60% of test data. From the result, we conclude the method has the predictive ability for new patient data. We are planning to use this simulator in clinical scene and verify the effectiveness of it.