Building　a　predictive　model　to　understand　a　patient＇s　status　and　the　future　progression　of　disease　is　an　important　precondition　for　facilitating　preventive　care　to　reduce　the　burden　of　chronic　obstructive　pulmonary　disease.　Recently,　unsupervised　feature　learning　has　become　the　dominant　approach　to　predictive　modeling　based　on　EMRs　(electronic　medical　records),　which　usually　adopts　word　and　concept　embeddings　as　distributed　representations　of　clinical　data.　By　observing　that　(1)　semantic　discrimination　is　limited　by　the　inherent　small　feature　granularity　of　words　and　concepts　and　(2)　clinical　decision-making　is　conducted　based　on　a　set　of　attribute-value　pairs,　e.g.,　a　clinical　laboratory　test　(the　attribute)　with　a　numerical　or　categorical　value,　this　paper　proposes　a　novel　predictive　modeling　approach　based　on　EMRs.　In　this　approach,　multi-granular　features,　i.e.,　words,　concepts,　concept　relations　and　attribute-value　pairs,　are　extracted　through　three　subtasks　(concept　recognition,　relation　extraction　and　attribute-value　pair　extraction),　then,　combined　to　derive　representations　and　predictive　models　of　chronic　obstructive　pulmonary　disease　(COPD).　The　approach　itself　is　highly　generic,　it　can　be　used　for　different　disease,　but　limited　in　Chinese　EMRs.　In　this　paper,　we　focus　on　the　COPD　risk　prediction,　and　conduct　extensive　experiments　on　real-world　datasets　for　a　comparison　study.　The　results　show　that　our　approach　outperforms　baselines,　which　demonstrates　the　effectiveness　of　the　proposed　model.