Purpose/Significance To build an effective automatic classification method for a large number of medical literatures, so as to cope with the new challenges brought by the rapid growth of medical literatures for their classification and utilization. Method/Process Taking medical literatures as data source, the study utilizes the synonyms and hierarchical structure of the medical subject headings (MeSH) to enhance the semantic features of concept information, uses bidirectional encoder representations from transformers (BERT) for fine-tuning and testing, and compares the classification results with random forest (RF).Result/Conclusion The results of the ten-fold cross-validation method show that the precision, recall and F1 score of this medical literature classification method are 95.42%,93.61%,94.47%,which are better than the classification results of RF and other methods without feature enhancement, and show accuracy, effectiveness and applicability.