学校代码：10004密级：公开北京交通大学
硕士专业学位论文
pheic基于真实世界临床数据的中医动态诊疗方案推荐方法研究
Rearch on Recommendation Method of Dynamic Diagnosis and Treatment Scheme of Traditional Chine Medicine
bad on Real World Clinical Data
作者姓名：何雄学号：18125206
导师姓名：周雪忠职称：教授
工程硕士专业领域：计算机技术学位级别：硕士
北京交通大学
2020年6月
致谢
英语招聘
时光匆匆，转眼间即将面临毕业，研究生生活时间虽短，却经历了太多美好。回想过去的这段时光，收获了太多的知识和幸福，感谢这一路来帮助我陪伴我的老师、同学和朋友，脑海中永远会记住你们这些可爱的人。虽然以后的路还很长，还会经历太多的未知和迷茫，但我会带着这段宝贵的回忆，继续勇往直前。
likeyou什么意思英文首先要感谢我的导师周雪忠老师。周老师严谨认真的科研态度令人折服，对学生非常地和蔼可亲，总是想着如何提升我们的能力和科研水平，给我们提供了一个高水平的平台，让我们得到充分锻炼和成长。非常幸运能成为周老师的学生，这两年真的学到了很多，每次跟周老师交流结束，都有醍醐灌顶的感觉，所有的困惑一下变得明朗，研究生期间不但专业能力和学术能力取得了很大的提升，而且还开阔了眼界，提升了表达能力和综合素质。
其次，我要感谢中医科学院文天才老师。入学前的暑假开始在数据中心实习，文老师带着我做了很多的项目，从最开始接触数据到后来深入地去研究模型，文老师一步一步带着我成长。在中医科学院数据中心最快乐的时光就是中午与文老师讨论交流，不仅解决了项目中遇到的问题，也解决了我生活和工作中的困惑。此外，我还要感谢潘溪水老师和张小平老师，感谢他们给予我的帮助和照顾。
接下来，我要感谢我的师兄、师姐和同学在研究生期间给予我的关心和帮助，感谢杨扩师兄、王宁师兄、胡晓晨师兄和满雨桐师姐在写论文期间给予的指导，感谢钟全、汤文明和邢颖在写论文上给予的
帮助。感谢实验室常老师、舒梓欣师姐、王晶晶师姐、邹群盛和闫佳伟在学习和生活中给予的关心和帮助。此外还要感谢我的室友韩佳成、洪文博、王振学给予生活上的陪伴、关心和照顾，因为他们，研究生期间增添了太多的幸福和快乐。
除此之外，我还要感谢一直以来关心支持我的父母和家人，感谢他们的默默付出，因为有他们坚强的后盾，我才能专注于我的学业，不断向前。
最后，我要感谢各位评审老师百忙之中抽出宝贵时间审阅我的论文，您的意见将有助于我的研究工作更进一步提升，敬请各位老师批评指正，各位老师辛苦了！
fack you摘要
医学人工智能成为当前研究热点之一，将人工智能相关技术应用于医疗领域，不仅能实现诊疗智能化，同时也能显著提升整体医疗水平，减少医疗成本。在中医领域，处方推荐是辅助诊疗的核心问题之一，由于中医处方药物组合的复杂性、患者合并疾病的复杂性以及患者的个性化差异，处方推荐依然是中医人工智能中最为困难的研究问题。处方推荐可以分为两种模式，一是学习医生积累的经验，通过症状来预测处方；二是将患者的治疗看作序贯决策过程，挖掘患者的治疗路径，预测最佳诊疗方案序列。随着大规模真实世界临床数据的积累和深度学习、强化学习技术发展，使得构建精准模型进行处方推荐成为可能。本文基于大规模真实世界临床数据，首先研究糖尿病合并疾病预测和处方预测，
中文菜单英文译法然后在此基础上，进一步研究患者动态诊疗方案优化问题，本文主要工作分为以下三个方面。
开房屋租赁发票
（1）针对临床患者的合并疾病预测问题，提出一种融合临床与多源网络数据的图卷积神经网络分类模型（GCN-CMSN）。首先我们利用临床患者的疾病合并关系以及多次住院信息，构建20000例的糖尿病合并疾病标准数据集。然后利用随机森林算法分析糖尿病的相关风险疾病。最后提出融合疾病基因关系以及临床疾病合并关系的糖尿病合并疾病预测模型（GCN-CMSN），不同于以往模型，GCN-CMSN能够融合临床与多源疾病网络数据，进行患者合并疾病的预测，实验结果表明，该模型在查全率和F1值上都取得最好的效果。top什么意思
ownskin（2）针对特定疾病人群的中医处方推荐研究，提出一种基于编码器和解码器的多任务多标签处方预测模型（MTL-ED）。首先对中国中医科学院数据中心提供的中医治疗糖尿病数据进行预处理，构建糖尿病症状处方数据集。然后提出一种面向多标签分类任务的算法MTL-ED，该算法通过患者的症状预测治则治法和中医处方，同时预测出每味药物的相关症状，最后将该处方推荐给患者或医生以供参考。实验结果显示，算法在准确率和召回率等评价指标上取得较大提升。
pws
（3）针对特定疾病人群的中医动态诊疗方案优化问题，提出基于深度强化学习的诊疗方案优化方法。首先构建糖尿病序贯诊疗标准数据集，然后构建基于深度强化学习的两阶段处方推荐框架，该框架第一阶段利用强化模型优化诊疗方案，第二阶段根据患者的症状和诊疗方案预测中药处方。此外，
设计了一种预测患者状态转移的虚拟环境。实验结果表明，强化模型推荐的诊疗方案使患者具有更佳的愈后，比医生给出的诊疗方案序列更加合理。
关键词：真实世界临床数据；图卷积神经网络；Seq2Seq；强化学习；动态诊疗方案；中医处方推荐；糖尿病
ABSTRACT
Medical artificial intelligence has become one of the current rearch hotspots.The application of artificial intelligence related technologies in the medical field can not only realize the intelligent diagnosis and treatment,but also significantly improve the overall medical level and reduce medical costs.In the field of TCM(Traditional Chine Medicine),prescription recommendation is one of the core issues of assisted diagnosis and treatment.Becau of the complexity of the combination of TCM prescription drugs, the complexity of patients with combined dias and the individual differences of patients,prescription recommendation is still the most difficult rearch problem in TCM artificial intelligence.Prescription recommendation can be divided into two modes:one is to learn the experience accumulated by doctors and predict the prescription through symptoms;the other is to regard the treatment of patients as a quential decision-making process,mine the treatm
ent path of patients and predict the quence of the best treatment plan.With the accumulation of large-scale real world clinical data and the development of deep learning and reinforcement learning technology,it is possible to build accurate models for prescription recommendation. Bad on the large-scale real world clinical data,this paper first studies the prediction of diabetes mellitus combined with dia and prescription,and then further studies the optimization of patients'dynamic diagnosis and treatment scheme.The main work of this paper is divided into the following three aspects.
夜猫子 英文(1)To solve the problem of dia prediction for clinical patients,A graph convolution neural network classification model bad on clinical and multi-source network data(GCN-CMSN)is propod.First of all,we construct a standard data t of 20000cas of diabetes mellitus combined dia by using the dia combination relationship of clinical patients and multiple hospitalization information.Then we u random forest algorithm to analyze diabetes related risk dias.In the end,a new model of diabetes mellitus combined dia prediction GCN-CMSN is put forward, which is different from the previous model.GCN-CMSN can integrate clinical and multi-source dia network data to predict patients combined dia.The experimental results show that the model has the best effect on recall and F1.
(2)A multi-task and multi-label prescription prediction model bad on encoder and decoder(MTL-ED)is propod for the study of TCM prescription recommendation
for specific dia population.Firstly,the data of diabetes treatment provided by the data center of the Chine Academy of traditional Chine Medicine was preprocesd to build the data t of diabetes symptom prescription.Then,an algorithm MTL-ED for multi-label classification task is propod.The algorithm predicts the treatment principle and the prescription of TCM through the symptoms of patients,and predicts the relevant symptoms of each drug.Finally,the prescription is recommended to patients or doctors for reference.The experimental results show that the algorithm has achieved a great improvement in Precision and Recall rate.
(3)Aiming at the optimization problem of dynamic diagnosis and treatment scheme of Chine medicine for specific dia population,the optimization method of diagnosis and treatment scheme bad on deep reinforcement learning is propod. Firstly,the data t of diabetes quential diagnosis and treatment standards is constructed,and then a two-stage prescription recommendation framework bad on deep reinforcement learning is constructed.In the first stage,the reinforcement model is ud to optimize the diagnosis and treatment scheme,and in the cond stage,the TCM prescription is predicted according to the symptoms and treatment scheme of
patients. In addition,a virtual environment is designed to predict the state transition of patients. The experimental results show that the diagnosis and treatment scheme recommended by the enhanced model makes the patients have better prognosis,and is more reasonable than the diagnosis and treatment scheme quence given by doctors. KEYWORDS：real world clinical data;convolution neural network;Seq2Seq; reinforcement learning;dynamic diagnosis and treatment program;traditional Chine medicine;diabetes;
