【作者】 李林峰；

【导师】 孙冬梅； 王升辉；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2020， 博士

【摘要】 临床决策支持系统(Clinical Decision Support System,CDSS)有着重要的社会价值。国家在《“健康中国2030”规划纲要》和《关于促进“互联网+医疗健康”发展的意见》中都明确要求研发基于人工智能的临床诊疗决策支持系统。CDSS是基于知识库和人工智能技术,为医生在疾病的诊断、治疗、风险预测、合理用药等方面提供决策支持的计算机系统。近年来,大数据和自然语言处理、知识图谱、机器学习等人工智能技术的飞速发展给CDSS带来了新的机遇和挑战。本文围绕CDSS中的人工智能关键技术做了以下几方面的工作:(1)提出了基于真实世界大规模电子病历构建医学知识图谱的方法本文提出了一套基于真实世界大规模电子病历构建医学知识图谱的系统流程。该流程包含七个步骤,针对关系抽取、属性计算、图谱清洗和相关实体排序四个步骤提出了新的方法;同时鉴于医学知识图谱的特殊性,本文提出了四元组知识点表示方法和综合考虑实体关系的发生概率、特异度及可信度的排序算法PSR(Probability Specificity Reliability)。该流程在基于百万级电子病历构建医学知识图谱中得到验证,所构建的知识图谱是高质量的,且包含的医学统计信息更加丰富。基于此方法构建的医学知识图谱已经成功的应用于西南医院全科临床决策支持系统中。(2)提出了医学概率知识图谱嵌入算法Pr Trans X由于医学知识图谱中实体间的关系不是确定的,而是具有一定的概率性,这有别于通用领域的知识图谱,因此传统的嵌入算法无法应用在医学概率知识图谱中。针对这一问题,本文提出了一系列基于概率转移的嵌入算法Pr Trans E/H/R/D/Sparse(简记为Pr Trans X),通过定义关系三元组的距离分数与概率值之间的映射函数,重新构建了包含正负三元组概率损失的目标函数,使得学习目标更符合医学知识图谱的真实情况。我们利用Pr Trans X算法对上述基于电子病历构建的医学概率知识图谱进行了嵌入学习,并通过链接预测和其他医学预测任务验证了,Pr Trans X算法的性能显著优于传统的Trans X算法,可以明显提升预测实体间关系的准确率。(3)构建了基于思维导图的可扩展计算的智能临床决策引擎传统的基于思维导图的临床决策引擎构建过程是割裂医学专家和计算机专家的,会因两个领域的理解不同导致决策引擎的推荐结果错误,或不被医学专家接受。因此,本文提出了一种新的基于思维导图的可扩展计算的智能临床决策引擎构建方法,医学专家可以和计算机专家共同定制规则决策树,且在医学计算引擎中内置了多个算子,可以通过配置计算表达式完成绝大部分工作。同时该引擎支持自定义算子,针对不同的任务,扩展配置不同的计算表达式,使得它具有很强的灵活性,为第五章实现规则和机器学习的混合模型搭建了基础。我们基于该引擎开发了肝癌CDSS,它支持肝癌患者14个重要决策场景的辅助决策,从开发过程和结果来看,该引擎在开发基于临床指南的CDSS上具有一定的技术先进性。(4)提出了规则决策树中嵌入机器学习的混合决策模型之前的研究工作表明,单纯基于数据的机器学习建模类CDSS准确率虽高,但其解释性差导致临床接受度不高;而基于临床指南的规则类CDSS准确率较差,但因其解释性好更容易被临床接受。针对该问题,本文提出了一种在规则决策树中嵌入机器学习的混合决策模型。具体来说,以临床指南转化后的规则为基本框架,将机器学习模型嵌入到规则决策树的中间节点和叶子节点,保证总体符合指南规范的前提下,在个别节点充分发挥数据建模的优势。通过将该方法应用于结肠癌术后辅助化疗的方案推荐上,发现混合模型的工作机制更接近临床医生理性(符合指南规范)与感性(临床经验)共同决策的真实过程,实验结果也验证它在准确性和可解释性上都优于其他方法。更多还原

【Abstract】 Clinical decision support system(CDSS)is important to society.The State Council has clearly emphasized the importance of research and development of CDSS based on artificial intelligence by issuing many policy documents.Based on clinical knowledge and artificial intelligence technology,CDSS provides decision support for doctors in diagnosis,treatment,risk prediction,rational medication of diseases and other aspects.In recent years,the rapid development of big data,natural language processing,knowledge graph,machine learning and other artificial intelligence technologies has brought new impetus to CDSS.In this dissertation,we give a detailed analysis of the overall structure and study the four core technologies of the CDSS.Our main works are summarized as follows,we(1)Constructed a systematic approach to build the medical knowledge graph(KG)from real-world electronic medical records(EMRs)This systematic approach is composed of seven steps,and we created new methods for four steps,respectively,relation extraction,property calculation,graph cleaning,and related-entity ranking.For relation extraction step,we proposed a novel quadruplet structure to represent medical knowledge instead of the classical triplet in KG.And a novel related-entity ranking function considering probability,specificity and reliability(PSR)is proposed.The established systematic approach can efficiently construct a highquality medical KG from large-scale EMRs.The proposed ranking function PSR achieves the best performance under all relations.Moreover,the obtained medical KG finds many successful applications due to its statistics-based quadruplet,such as the CDSS of Southwest Hospital.(2)Proposed a series of algorithms named Pr Trans X to embed the medical probabilistic knowledge graphIn medical knowledge graphs,the relations between head and tail entities are inherently probabilistic,that is the main difference from typical knowledge graphs in which the relations are deterministic.To address the challenge,we enhanced the existing Trans X(X=E/H/R/D/Sparse)algorithms to Pr Trans X by 1)constructing a mapping function between the score value and the probability and 2)introducing probability-based loss of triplets into the original margin-based loss function.The proposed Pr Trans X algorithms are performed on the medical KG obtained by above approach.The disease clustering result validates the efficacy of the learned embedding vector as entity’s semantic representation.Then,the embeddings are evaluated using link prediction and another medical prediction task,the Pr Trans X performs better than the corresponding Trans X model.(3)Designed an intelligent and scalable computing clinical decision engine based on mind mappingBased on such engine,medical experts and computer experts can customize the rule decision tree together.And such engine has built-in multiple medical calculation operators,that means most of the work can be done by configuring calculation expressions.In addition,the designed decision tree can support user-defined operators,which makes it very flexible,and builds the foundation for the later hybrid model of rule decision trees and machine learning models.We developed a CDSS of liver cancer based on the proposed intelligent clinical decision engine.This system supports the 14 important decision scenarios of liver cancer patients.And the results show that the engine is suitable for developing CDSS system based on clinical guidelines,and has certain advanced technology in this field.(4)Built a hybrid decision model by embedding machine learning models into rule decision treeThe accuracy of data-based CDSS is high,but its clinical acceptance is not high because of its poor interpretation and sometimes may not conform to the clinical guidelines.On the other side,the accuracy of knowledge-based CDSS is poor,but it is accepted more easily because of its good interpretation.To solve these problems,we built a hybrid decision model,which can embed the machine learning model into the middle and/or leaf nodes of the rule decision tree.It makes the best of the advantages of data modeling in the individual nodes on the premise that the whole model conforms to the guidelines.We applied this hybrid model to the recommendation of adjuvant therapy for colon cancer.Compared with the simple rule-based or simple machine learning model,the working mechanism of the hybrid model is closer to the real process of rational(in line with the guidelines)and perceptual(clinical experience)decision-making of clinicians.The experimental results also verified that it is superior to other methods in accuracy and interpretability.更多还原