Research On Biomedical Information Extraction In Chinese EMRs Based On Deep Learning Methods

EMRs,as the medical resource constructed by professionals,contain a great size of medical data including various digital information like words,tables and images.There are also a great amount of medical knowledge hided in these records.Analyzing and mining these knowledge can not only allow us to have a clearer understanding of the relation between pathological symptoms and diagnosis,but also to dig deeper pathological relationships and finally achieve the purpose of supporting diagnosis.In this way,it has become a primary and key process to extract the concepts in EMRs.In this paper,we firstly sorts out the research status of the field of Chinese concept extraction in EMRs,and then summarizes the main problems in the current field.At the same time,inspired by the excellent performance of deep learning in various fields,I introduce the deep learning methods into this field.The traditional framework of deep learning dealing with NLP tasks is adopted,namely word embedding training and RNN model training,and various word embedding training and RNN based models are experimented.Besides,regarding to the linguistic feature of Chinese,two new algorithms,Chinese biomedical information extraction algorithm basing on character embeddings and Chinese Biomedical information extraction algorithm basing on character and word united embeddings are adopted.The experiments in the following chapters also validate the effectiveness of the two algorithms.Here is the main content of the paper:(1)Annotated corpus construction and data preprocessingTwo concepts,position and symptom,are defined in this paper.The concept annotation work is conducted by a team of professionals.Then a series of preprocessing procedures including annotation content extraction,sentence and word segmentation,corpus construction,sentence length pruning.(2)Introduction of deep learning methods into the field of Chinese Biomedical information extraction fieldThe traditional framework of deep learning dealing with NLP tasks is adopted,namely word embedding training and RNN model training.The word embedding model are trained on the combined data set of annotated and unannotated data.Then the trained word embedding initializes the training of various RNN based model.(3)Two new algorithms specified for Chinese Biomedical information extractionIn the English research field,usually only word embedding is considered when build the deep neural networks.Regarding to the linguistic feature of Chinese,two new algorithms,Chinese Biomedical information extraction algorithm basing on character embeddings and Chinese Biomedical information extraction algorithm basing on character and word united embeddings are adopted.And various experimental results in the following chapters also validate the effectiveness of the proposed algorithms.