[Purpose/significance] This study aims to explore the significance of factors on the rumor identification in the social media environment and identify rumors in public health emergencies. We also evaluated the important features that affect the identification of rumors to help the cyber security department accurately identify rumors and maintain a healthy network ecological environment. [Method/process] We extracted user features, time features, structure features, text semantic features and propagation features in microblog entries. We combined with the MAIN theoretical models, and used binary logistic regression method to deeply research the influence factors of rumors from the perspective of modality, information content, information sources. We built a multi-feature based rumor identification model that integrated the semantic feature extracted by neural network model. XGBoost algorithm was used to calculate the importance of different features in rumor identification. [Result/conclusion] The higher the positive emotional value of comment, the number of microblog entries posted by users, and greater the influence of users, the lower the possibility that the microblog entry is a rumor. The value of the accuracy of rumor recognition model is 0.984. The semantic features of text are the most important.