Electronic Medical Records (EMRs) contain a lot of information about patients’ medication history. Such information can be used for clinical research to discover the risk of complications from diseases, especially for long-term treatment of chronic diseases such as Type 2 Diabetes Mellitus (T2DM). In T2DM, if the symptoms are detected earlier, these complications can be prevented or delayed. Screening process by manual chart review takes a lot of time, some patients might be overlooked and patients’ characteristics might change along with age. To improve the screening process, we propose a method by temporal analysis of patients’ phenotypes to predict risk of complication using EMRs data. The purpose of this research is to identify the feasibility of predicting risk of diabetic nephropathy from long-term treatment of T2DM by using temporal phenotyping. In this approach, we use structured EMRs data of T2DM patients from Kyoto University Hospital. To observe the change in the phenotypes, we divide patients' medication history from prescription data into episodes by reconstructing medication history according to the notion of stable period. We apply rule-based algorithm to each medication episode for classifying patients, according to phenotypes associated with diabetic nephropathy and apply machine learning method to predict the possible outcome of future phenotypes.