Current　sudden　cardiac　death　(SCD)　studies　mostly　use　traditional　machine　learning　algorithms　and　suffer　from　low　accuracy.　Deep　learning　has　a　promising　application　in　the　field　of　SCD　research.　The　study　extract　R-R　interval　and　R　amplitude　from　ECG　signals　as　inputs,　combined　convolutional　neural　network　and　gated　recurrent　unit,　and　take　full　advantage　of　hybrid　neural　network　structure　to　realize　the　risk　stratification　of　high-risk　patients　who　may　have　SCD　within　90　minutes,　with　the　highest　accuracy　of　95.33%.