This　paper　studies　the　optimal　feature　subset　screening　for　diabetes　according　to　the　health　check　data　based　on　the　random　forest　algorithm.　The　paper　takes　the　real　physical　examination　records　of　the　same　batch　of　people　in　a　local　health　check-up　center　from　2010　to　2015　as　the　data　source,　and　evaluates　the　importance　of　the　features.　The　preliminary　fitting　finds　that　28　features　have　an　impact　on　the　response　results.　The　AUC　performance　of　the　classifier　finally　selects　the　optimal　feature　subset　containing　9　characteristic　variables　in　multiple　feature　subsets,　which　provides　scientific　evidence　and　decision　support　for　medical　expert＇s　prediction　intervention,　clinical　diagnosis,　treatment　plan　determination　and　medical　research　on　diabetes.　©　2019　IOP　Publishing　Ltd.　All　rights　reserved.