With　the　increasing　interests　of　second　use　of　medical　data,　concept　extraction　in　Electronic　Medical　Records　has　drawn　more　and　more　scholars＇　attention.　Owing　to　the　artificial　data　annotation　task　is　labor　intensive,　the　method　of　concept　extraction　is　mainly　to　use　the　fully　labeled　documents　as　training　data　in　order　to　build　a　concept　instance　identifier.　However,　in　many　cases,　the　available　training　data　are　sparse　labeling.　This　fact　makes　the　performance　of　the　constructed　classifier　is　poor.　Existing　methods　for　extracting　concepts　either　considered　the　diversity　of　datasets　or　considered　the　various　learning　models.　Therefore,　this　paper　proposes　a　novel　approach　to　improve　the　performance　of　concept　extraction　from　electronic　medical　records　by　combining　the　diversity　of　datasets　with　the　various　learning　models.　The　large　sparsely　labeled　dataset　is　split　into　multiple　subsets.　Then　the　different　subsets　are　trained　by　different　learning　models,　such　as　HMM,　MEMM,　and　CRF,　in　an　iterative　way.　Our　technique　leverages　off　the　fact　that　different　learning　algorithms　have　different　inductive　biases　and　that　better　predictions　can　be　made　by　the　voted　majority.　©　2016　ACM.