With the increasing number of diabetes patients, the developments of continuous glucose monitoring (CGM) techniques and glucose prediction models become increasingly important. In this study, we propose a personalized blood glucose prediction approach via glycemic oscillation decomposition based on CGM data. We first utilize advanced data analytics techniques to decompose CGM data into multiple patterns through the oscillation pattern mining module. The temporal pattern learning module is then developed to capture the temporal dependency of glucose levels. We further aggregate multiple outputs into glucose predictions with improved accuracy. We conduct a comparison study between the proposed approach and other existing models using OhioT1DM dataset. Experimental results show that the proposed work can provide more accurate predictions for diabetic glucose levels compared to other methods. With improved prediction performance, the proposed approach facilitates personalized blood glucose management services for diabetes patients.