We analyzed the contents and structure of current electronics medical records,and proposed a definition of Five-Tuples pattern and another more fine-grained definition of two-turples pattern and semantic classes.On this foundation,we proposed a series of algorithms including patterns generalization,patterns automatic extraction and medical information extraction.The experiments with 312 actual medical records show that the system performs well both in the precision and recall.And because of the functionality of self-learning,the system will be more outstanding with an increase in the training corpus.