A method for generating representations of temporal clinical events is proposed. The proposed method captures the sequential information of longitudinal EHR data. Using learned sequences as features outperforms single temporal representation. Learned sequences lead to more informative features than domain knowledge.