Abstract: 【Purpose/significance】 This research aims to explore the thematic features of real and false information, study the problem of identifying the authenticity of comment information, and provide reference basis for information recognition on social media platform under the background of public health events.【 Method/process】 For epidemic related multi topic data on social media platforms, LDA models are used to extract thematic features of real and false information. By introducing a BERT preprocessing method, we construct a BERT-BiLSTM hybrid model to identify false epidemic information.【 Result/conclusion】 We found that there are significant differ⁃ ences between real and false information in theme features and expression methods, providing opinions and references for identifying false information. In addition, compared with traditional machine learning algorithms, BERT-BiLSTM model has significant advan⁃ tages in identifying epidemic misinformation, with an accuracy rate of 0.960 and an F1 value of 0.961. The BERT-BiLSTM model will provide a more efficient and accurate solution for misinformation recognition.【 Innovation/limitation】 Taking epidemic information on social media platforms as the research object, the LDA model was comprehensively used to explore the main characteristics of real and false epidemic information. Effective identification of multi topic data was achieved at a lower cost on small-scale datasets, providing an efficient solution for infodemic management.