Extracting Entities With Attributes In Clinical Text Via Joint Deep Learning

With the rapid development of information technology,clinical medical services have entered the era of digitalization and informationization.The establishment and popularization of a series of medical information systems with electronic medical records as the core has laid a solid foundation for medical data.In the era of medical big data,how to analyze and utilize medical big data to improve the intelligence of clinical medical service has become one of the problems that need to be solved urgently.Among them,the clinical medical entity and its attribute extraction is the key to the medical information mining of electronic medical records,and the key steps of medical knowledge mining.The clinical medical entity and its attribute extraction is to identify the clinical medical entities and attributes in medical texts,and to determine the modification relationship between the entities and attributes,which can effectively organize the rich medical information and knowledge hidden in the electronic medical record.Clinical medical entities and their attributes provide the basis for data analysis and mining for patients,medical staff,and researchers,and are widely used in clinical decision support systems,personalized health care information services,and public health services.Traditionally,the task of clinical medical entities and their attribute extraction is the pipeline method.This approach makes each sub-task a separate model,simple and easy to implement,however it ignores the intrinsic relationship and dependencies between the two sub-tasks,and inevitably causing error propagation.The joint learning method uses a single unified model to directly identify the clinical medical entities,attributes and their modification relationship from the medical text,which can effectively integrate the internal information between the clinical medical entities and its attributes,and improve the recognition rate of entities and attributes through the correct relationship extraction.At the same time,since the correct modification relationship cannot be obtained from the wrong clinical entities and attributes,improving the recognition rate of entities and attributes will also improve the extraction performance of the modification relationship between entities and its attributes.In this paper,the following two methods are used to jointly extract clinical medical entities and their attributes:1)Serial joint method:clinical medical entities and attribute recognition are treated as sequence labeling problems,and the relationship between them is treated as a classification problem;2)Parallel joint method:The clinical medical entity and its attributes are both extracted as a sequence labeling problem.The former uses an end-to-end neural network framework to jointly extract medical entities and their attributes by sharing the parameters of two subtasks.Based on the existing research,this paper proposes two new serial joint deep learning methods for extracting medical entities and their attributes.These methods are further improved by introducing clinical medical constraint relationship and bias loss function.The latter is to transform two sub-tasks into one task by designing a new problem representation.In this paper,BIOHD1234 and Multi-Label are proposed for parallel joint learning of medical entities and their attributes.In order to evaluate the proposed method,a Chinese electronic medical record data set is constructed and tested on Chinese data set and English SemEval-2015 Task 14 public corpus respectively.The experimental results show that the new joint learning method proposed in this paper is better than the traditional pipeline method and the joint learning method proposed by Miwa and Bansal^[1].The joint learning methods perform better pipeline methods.And the serial joint learning methods perform better than the parallel joint learning methods.To the best of our knowledge,this is the first study to investigate the joint task of clinical entity/attribute recognition and clinical-attribute relation extraction in Chinese clinical text by using a deep learning framework.