Due　to　its　continuously　increasing　occurrence,　more　and　more　families　are　influenced　by　diabetes　mellitus.　Most　diabetics　know　little　about　their　health　quality　or　the　risk　factors　they　face　prior　to　diagnosis.　In　this　study,　we　have　proposed　a　novel　model　based　on　data　mining　techniques　for　predicting　type　2　diabetes　mellitus　(T2DM).　The　main　problems　that　we　are　trying　to　solve　are　to　improve　the　accuracy　of　the　prediction　model,　and　to　make　the　model　adaptive　to　more　than　one　dataset.　Based　on　a　series　of　preprocessing　procedures,　the　model　is　comprised　of　two　parts,　the　improved　K-means　algorithm　and　the　logistic　regression　algorithm.　The　Pima　Indians　Diabetes　Dataset　and　the　Waikato　Environment　for　Knowledge　Analysis　toolkit　were　utilized　to　compare　our　results　with　the　results　from　other　researchers.　The　conclusion　shows　that　the　model　attained　a　3.04%　higher　accuracy　of　prediction　than　those　of　other　researchers.　Moreover,　our　model　ensures　that　the　dataset　quality　is　sufficient.　To　further　evaluate　the　performance　of　our　model,　we　applied　it　to　two　other　diabetes　datasets.　Both　experiments’　results　show　good　performance.　As　a　result,　the　model　is　shown　to　be　useful　for　the　realistic　health　management　of　diabetes.　©　2017