Diabetes　mellitus　is　a　metabolic　disorder　caused　by　the　absolute　insufficient　secretion　of　insulin　and　the　disorder　of　insulin　utilization.　Diabetes　mellitus　will　bring　great　harm　to　the　organs,　and　the　complications　of　diabetes　will　pose　a　great　threat　to　the　health　and　life　of　patients,　and　even　lead　to　disability　and　death.　The　prediction　of　diabetes　has　always　been　a　hot　topic,　but　it　is　very　difficult　to　predict.　From　a　medical　point　of　view,　in　this　study,　we　aim　to　establish　a　diabetes　prediction　model　based　on　machine　learning　and　data　mining.　We　first　proposed　a　dual　characteristic　variable　selection　method　based　on　single-factor　regression　and　LightGBM,　which　can　screen　out　the　medical　indicators　affecting　diabetes.　On　this　basis,　we　built　a　single　diabetes　prediction　model　based　on　machine　learning,　and　further　studied　XGBoost　and　ResNet.　Finally,　we　used　GA2Ms,　XGBoost　and　ResNet　to　study　the　diabetes　prediction　model　based　on　ensemble　learning.　The　results　show　that　the　accuracy,　F1　and　AUC　of　the　prediction　model　are　0.853,　0.888　and　0.875　respectively　after　five-fold　cross-validation　and　comparative　analysis,　which　are　significantly　better　than　other　machine　learning　models.　Therefore,　the　proposed　method　can　accurately　predict　diabetes,　so　as　to　provide　effective　clinical　auxiliary　diagnosis　for　doctors,　help　doctors　take　preventive　measures　in　advance,　improve　the　survival　rate　of　patients,　and　reduce　the　impact　of　diabetes　on　patients.　©　2021　IEEE.