Research On Decision Rule Extraction For Medical Text

A clinical decision support system(CDSS)is a typical application of artificial intelligence in the medical field,assisting medical practitioners in making decisions,medical teaching,and advising patients.However,as the core of CDSS,the construction of decision rules relies on manual annotation by medical experts,which is time-consuming and cannot absorb the latest research timely.All these hinder the construction,dissemination,and maintenance of large-scale CDSS.To automate the extraction of medical decision rules and provide knowledge support for intelligent healthcare,we propose a novel information extraction task,Text2 MDT,to explore the automatic extraction of medical decision trees with decision rules modeled(MDTs)from medical texts.The MDT is composed of condition nodes and decision nodes,which represent the process of making medical decisions based on the results of different conditions of patients.It is a structured and standardized representation of medical decision rules.In this paper,we constructed the first text-to-MDT dataset with medical experts’ participation and experimented with benchmarking methods based on this dataset,laying the foundation for the automatic extraction of medical decision rules and the automated construction of large CDSS in the future.Based on Text2 MDT,this paper proposes a Tabel-Filling based decision tree extraction method(T2T)to automatically extract MDTs.The method is a two-stage approach:firstly,conditions and decisions are extracted from the text;then,a bi-affine model is used to predict the position of all conditions and decisions in the MDT,and an approximate decoding algorithm is used to generate the MDT based on the position relationship.Compared to traditional tree generation methods that use a sequential model to generate tree nodes at each time step,T2 T has no dependency on nodes in the tree generation phase,alleviating the exposure bias associated with traditional tree generation methods and obtaining a significant improvement in all evaluation metrics.In addition,as multi-stage learning methods can lead to error cascades and neglect the connection between the two stages,this paper propose a decision tree extraction method based on multi-turn question answering(Mt QA)to naturally fuse the two phases.Mt QA uses multiple questions to obtain the information needed for MDT generation.The process of answering the questions generates the MDT.It is demonstrated that Mt QA captures the potential connections between node semantics and tree structure and improves several metrics compared to the above multi-stage tree generation methods.