【作者】 王宁；

【导师】 余玉刚； 谢金贵；

【作者基本信息】 中国科学技术大学 ， 管理科学与工程， 2021， 博士

【摘要】 近年来,人工智能成为推动社会发展的新动力,正在重塑各个行业的运营模式。人工智能在医疗运作管理中的主要应用之一是开发决策支持工具,以优化临床诊疗决策,为患者提供更准确的诊断和治疗方案,提升医疗服务质量和效率,降低医疗错误和成本。信息化系统的广泛应用使得医疗机构积累了以电子病历为代表的海量医疗数据,通过挖掘电子病历数据隐藏的知识和信息,可以为临床诊断和治疗提供决策支持。中医作为我国最古老的传统医学,被应用于防治多种疾病,是我国医疗服务体系的重要组成部分。目前传统中医诊疗决策多依赖于医生专业能力和临床经验,主观性较强,诊疗水平参差不齐,并且缺乏信息化的临床辅助决策工具。中医电子病历不仅包含结构化数据,还包含隐藏大量有用信息的非结构化文本。然而,非结构化文本通常含有大量以古汉语表达的医疗专业术语,句法结构特殊且语义复杂,如何从中提取有用信息优化诊疗决策是一项重大的挑战。因此,基于中医电子病历的临床决策支持研究有重要的现实意义和理论价值。基于此背景,本文以通过数据驱动的方法优化诊疗决策为目标,基于电子病历中的结构化、非结构化数据,利用传统机器学习、深度学习和自然语言处理等方法,分别对中医证型预测、处方自动生成以及药物疗效相关性学习等中医临床诊疗中的三个关键性问题进行了研究,构建了从诊断、处方到疗效评价的中医临床决策支持体系。本文主要研究内容和结论如下:首先,针对临床诊断决策任务,本文第三章研究了基于预训练语言模型的中医证型预测,解决了如何辅助中医诊断病人的问题。本研究将中医证型预测建模为文本分类问题,直接以临床非结构化文本作为输入,提出了一种基于多个预训练语言模型的集成学习方法,构建了中医证型预测的端到端模型。实验结果显示本研究方法能够有效预测中医证型,模型评估指标宏平均F1和准确率分别为90.97%和94.23%。相比文献中主要的文本分类模型,本研究方法预测性能更优,有助于优化临床诊断决策。其次,针对临床治疗决策任务,第四章研究了基于序列到序列学习的中药处方自动生成,解决了如何辅助中医为病人制定处方用药的问题。本研究将中药处方自动生成任务建模为序列到序列的文本生成问题,提出一种采用特殊掩码机制微调预训练语言模型的方法,使得原本不具备序列生成机制的预训练语言模型可以完成中药处方生成任务。我们还利用药名全称编码对词级进行建模,进一步提高了模型的语义表示能力。实验结果显示本研究方法在生成高频中药上实现了较好的效果,模型评估指标召回率和精确率分别为78.24%和70.60%,处方生成表现明显优于相关文献中的模型。最后,针对药物疗效评价任务,第五章研究了基于关联规则的患者用药和治疗效果相关性学习,解决了如何辅助临床中医分析药物疗效的问题。本研究提出了基于关联规则方法的处方用药与治疗效果相关性学习框架。首先,我们分别定义了实验室单个指标是否有效改善和患者整体治疗结果是否有效改善,接着利用关联规则学习挖掘了处方用药与有效改善指标或整体治疗结果的关联模式,最后利用卡方检验提取统计学显著的关联规则。我们从单药物-单指标、单药物-多指标、多药物-单指标和多药物-多指标四个角度进行挖掘,并对比了不同治疗效果评价方法下的挖掘结果。实验结果得到临床医生和相关文献的验证,表明了该方法的有效性和可靠性。本文的主要贡献有:(1)本文对中医电子病历多类型数据进行了挖掘,特别是对非结构化文本进行了分析和建模,通过研究中医证型诊断、处方自动生成和中药疗效评价三个重要问题,进一步补充了中医临床诊疗决策支持研究。(2)本文为中医电子病历挖掘提供了新的研究方法。针对中医证型预测和处方生成问题,本文提出利用预训练语言模型在小规模数据训练的方法,建立了基于深度学习的端到端模型。针对疗效评价问题,提出了一种结合关联规则和卡方检验挖掘高维稀疏变量间的统计学显著关联规则的方法。(3)本文研究对提高中医临床诊疗决策质量有积极意义。通过构建从诊断、用药到疗效评价的中医临床决策体系,有效满足了临床实际需求。一方面可以为医生诊疗决策提供有益的参考,减轻医生负担并提高诊疗水平。另一方面,有助于增强患者诊疗的准确性和针对性,从而提升患者治疗效果并降低患者医疗费用。更多还原

【Abstract】 In recent years,artificial intelligence has become a new driving force to promote social development,and is reshaping the operations across industries.One of the main applications of artificial intelligence in healthcare operations is to develop decision sup-port tools,which help to optimize clinical diagnosis and treatment decisions,provide more accurate diagnosis and treatment plans for patients,improve the quality and effi-ciency of healthcare services,and reduce medical errors and costs.With the wide appli-cation of information systems,medical institutions have accumulated massive medical data represented by electronic medical records(EMR).By mining the hidden knowl-edge and information in electronic medical records,it can provide decision support for clinical diagnosis and treatment.As the oldest traditional medicine in China,traditional Chinese medicine(TCM)has been used in the prevention and treatment of various dis-eases,and is an important part of healthcare in China.At present,the diagnosis and treatment decision of TCM mostly relies on the professional ability and clinical expe-rience of doctors,which is highly subjective,uneven in diagnosis and treatment levels,and lacks information-based clinical decision-making tools.Electronic medical records not only have structured data,but also include unstructured texts that hide a large number of useful information.However,the unstructured text usually contains a large amount of medical terms expressed in ancient Chinese with special syntactic structure and complex semantics,and it is a major challenge to extract useful information from it to optimize diagnosis and treatment decision-making.Therefore,the study of clinical decision sup-port based on TCM electronic medical records has important practical significance and theoretical value.Based on this background,this thesis aims to optimize diagnosis and treatment decision-making by data-driven method.Based on structured and unstructured data in electronic medical records,we use traditional machine learning,deep learning and nat-ural language processing methods to study the three key issues in clinical diagnosis and treatment of traditional Chinese medicine,such as the prediction of TCM syndromes,the automatic generation of prescription and the association learning of Chinese herbal medicine and the treatment efficacy.Finally,a clinical decision support system of tra-ditional Chinese medicine from diagnosis,prescription to efficacy evaluation is con-structed.The main contents and conclusions of this thesis are as follows:First,for the task of clinical diagnosis decision-making,Chapter 3 of this thesis explores the prediction of TCM syndromes based on the pre-training language model,which solves the problem of how to assist TCM diagnosis of patients.Our study mod-els the prediction of TCM syndromes as a text classification and directly uses clinical unstructured text as input.We propose an integrated learning approach based on mul-tiple pre-training language models to construct an end-to-end model for the prediction of TCM syndromes.The experimental results show that the method is effective in the prediction of TCM syndromes,with model evaluation metrics of macro-average F1 and Accuracy of 90.97%and 94.23%respectively.Compared with the main text classi-fication models in the literature,our method has significant advantages in optimizing clinical diagnosis decision-making.Second,for the task of clinical treatment decision-making,Chapter 4 of the thesis studies the automatic generation of TCM prescription based on sequence to sequence learning,which solves the problem of how to assist TCM to formulate prescription.Our study models the automatic generation of TCM prescription as a sequence-to-sequence text generation problem.We propose a method to fine-tune the pre-training language model using a special masking mechanism,which enables the pre-training language model without the mechanism of sequence generation to complete the prescription gen-eration,We also model the word level using the full name encoding of herbs,which further improves the semantic representation of the model.The experimental results show that the method achieves good performance in the generation of high-frequency Chinese medicine,with the model evaluation metrics Recall and Precision of 78.24%and 70.60%respectively.And the prescription generation of our model significantly outperforms that of models in related literature.Finally,for the task of herbal efficacy evaluation,Chapter 5 of the thesis examines the correlation between Chinese herbal medicines and the treatment efficacy based on association rules,which solves the problem of how to assist clinical TCM analysis of herbal efficacy.Our study proposes a learning framework based on association rules for the correlation between Chinese herbal medicines and the treatment efficacy.Firstly,we respectively define whether a single laboratory index is effectively improved and whether the treatment overall outcome of patients are effectively improved.Then,we use association rule learning to mine the association patterns between Chinese herbal medicines and effectively improved indexes or the treatment overall outcome.Finally,we use chi-square test to extract statistically significant association rules.We mine the rules from four aspects:single herb-single index,single herb-multi indexes,multi herbs-single index and multi herbs-multi indexes.And we compare the mining results under different evaluation methods of treatment efficacy.Our experimental results have been verified by clinicians and related literature,indicating the effectiveness and reliability of the method.The main contributions of this thesis are as follows:(1)Our work mines multiple types of data from TCM electronic medical records,especially the unstructured text.Through the study of three important problems,i.e.the diagnosis of TCM syndrome,the automatic generation of the prescription and the evaluation TCM efficacy,the research on decision support of TCM clinical diagnosis and treatment is further supplemented.(2)Our work provides a new research method for mining the TCM electronic medical records.For the prediction of TCM syndrome and the generation of prescription,we propose a method of using pre-training language model to train in small-scale data,and establishe an end-to-end model based on deep learning.For the evaluation TCM ef-ficacy,a method combining association rules and chi-square test is proposed to mine statistically significant association rules among high-dimensional sparse variables.(3)Our work has a positive significance to improve the quality of clinical diagnosis and treatment decision-making of traditional Chinese medicine.Through the construction of TCM clinical decision-making system from diagnosis,medication to efficacy evalu-ation,it can effectively meet the actual clinical needs.On the one hand,it can provide a useful reference for doctors to make diagnosis and treatment decisions,reduce the bur-den of doctors and improve the level of diagnosis and treatment.On the other hand,it helps to enhance the accuracy and pertinence of patients’ diagnosis and treatment,so as to improve the treatment effect of patients and reduce the medical expenses of patients.更多还原