Myocardial infarction is a vascular necrosis of myocardial cells. Because of the different degrees of necrosis, the performance on the electrocardiogram is different. Because the characteristics of atypical myocardial infarction are not obvious, the current method has low detection accuracy. In this study, a method combining BP neural network and convolutional neural network is proposed according to the type of disease to improve the accuracy of detecting atypical myocardial infarction and reduce the time complexity of typical myocardial infarction. First, the original ECG signal is filtered, smoothed, de-baselined, etc., and the relevant feature values are extracted. The BP neural network is used to train the classification. For the features that are not obvious, the convolutional neural network is used for training after intercepting at a fixed length. In order to further improve the accuracy, the 12-lead data was combined for analysis. The experimental results show that this method improves the detection of myocardial infarction by 2% compared with other methods, especially in the detection of atypical myocardial infarction.