当前位置: 首页>博士论文>资源详情
诊疗流程差异检测与相似性搜索研究
中文摘要

“健康中国”战略已列入国家十三五规划,而伴随着的是基本医疗保险收不抵支的短板。同时,参保人得不到应有的服务和医保基金的骗保、浪费并存。为此,各地医保部门相继建立了基本医疗保险医疗服务监控信息系统,对接医院HIS系统开展具体诊疗服务监管,通过规范诊疗服务实现合理控费。这些监控信息系统根据医保部门监控人员预先定义的监控规则,筛查诊疗数据,生成违规疑点信息进行监控。预先定义的监控规则一般是单项监控规则,存在规则点离散、规则设置被动、规则数量有限使风险监控面窄以及规则内容独立等监控碎片化问题。规则间缺乏相互关联因而降低了监控的系统化,流程上没有覆盖全部节点因而难以实现全面监控。对于现代诊疗服务链条长、流转环节多、变化快的特点,存在单项监控规则自动发现新违规行为能力弱的问题。此外,若采用流程方式监控诊疗服务行为,面临选择和识别合理流程困难的问题;以及前期研究阶段真实诊疗流程数据获取困难问题。针对上述问题,本文研究诊疗流程差异检测、相似性搜索和流程模型生成方法,主要研究工作和创新如下: (1)针对目前各地基本医疗保险医疗服务监控信息系统的监控碎片化问题,本文提出基于特征行为的诊疗流程差异检测方法。 基于特征行为的诊疗流程差异检测方法,是将诊疗服务流程从行为上视为一组活动执行的序列,在任意流程模型对中自动识别行为差异。方法首先将给定的两个诊疗流程模型拆分成小的特征,模型特征在本文里指的是最常见的工作流模式,即顺序、分裂、合并以及循环。然后找出两个流程模型中相互匹配的活动,识别出和活动相关的差异模式。并找出存在匹配可能的特征,即两个特征中包含匹配的活动,识别匹配特征的差异模式。识别行为差异通过先分析特征的行为日志,再比较行为日志来分析行为差异,即用特征的依赖和痕迹来检测行为差异。最后返回两个流程模型间所有识别出来的行为差异。对于与合理流程行为差异较大的诊疗流程,可以认定为服务缺失或过度医疗,甚至违规、骗保直至欺诈。 (2)针对传统临床路径重指导性而与实际诊疗流程差距较大、以及具体个案诊疗流程普适性弱,缺乏选择和识别合理诊疗流程的方法问题,本文提出基于特征结构的诊疗流程相似性搜索方法。 基于特征结构的诊疗流程相似性搜索方法,是将诊疗服务流程从结构上视为图。本文以经典诊疗流程为基础,与同病种诊疗流程进行基于结构的灵活特征匹配,得到两个流程模型中的结构差异模式列表,然后计算每一种识别出来的差异特征间相似性的大小,根据所有差异特征相似性计算出两个流程模型间整体的相似性。从而搜索出相似度较高的诊疗流程群,形成相对合理区间的合理诊疗流程,或称合意流程组。并以该组作为流程差异比较的基准,对日后的流程进行合理性判断。本文的流程相似性搜索方法,是流程研究领域里第一个计算流程模型相似性并能够同时返回差异列表的方法。 (3)针对诊疗流程真实数据获取困难问题,本文提出基于现有真实诊疗业务流程模型特征自动生成特征类似的诊疗业务流程模型方法。 该方法首先分析给定诊疗流程模型集合的特征,特征主要包括标签特征(标签中的词频,词和词共同出现概率)、结构特征(顺序、分支等出现概率)、类型特征(活动和弧等的类型出现概率)。然后根据分析的特征生成模型片段,包括标签、结构等。再根据前面分析的概率向一个空模型中持续插入生成的片段,直到流程模型的大小达到预期要求。最后生成预期数量的诊疗流程模型。 (4)针对上述三种方法,本文利用部分真实诊疗流程及大量合成的诊疗流程,分别进行评估实验。针对流程差异检测方法的实验结果表明,每对流程的行为差异检测平均时间仅需0.011秒,且每个检测出的差异都与分解的小特征精确相关。针对流程相似性搜索方法的实验结果表明,千对流程的相似性计算平均时间为3.1秒,搜索出的相似流程有超过80%的准确率。针对生成合成流程方法的实验结果表明,真实流程中所有节点及边等元素几乎全部出现在2分42秒生成的1.03万个合成流程中,且合成流程与真实流程具有显著相似属性。综上所述,本文方法可以应用于对诊疗服务流程进行全面、高效监控。 关键词:诊疗流程差异检测;诊疗流程相似性搜索;合成流程模型;规范诊疗流程;医保医疗服务流程监控

英文摘要

"Health Chinese" has been included in the national strategic planning for 13th Five-Year, which is accompanied by the lack of funds for the basic medical insurance. While the insured donot get deserved service, and frauds and wastes coexist. To this end, medical insurance departments have established basic medical insurance and service monitoring information system. These systems, connecting HIS systems in hospitals, supervise medical services and control fees through standard rules. These monitoring information systems screen data for diagnosis and treatment and generate information about violations of suspected information monitoring, according to monitoring rules pre-defined by staffs of the health care department. Generally, the predefined monitoring rules are individual monitoring rules, such as discrete rule points, passive rules setting, limited number of rules, which raise the issue of monitoring fragmentation. The lack of correlation between the rules reduces the systematization of monitoring, and does not cover all nodes in the process, so it is difficult to achieve overall monitoring. Morden treatment services have a long service chains and change rapidly, which requires high adaptive ability. In addition, if the process is used to monitor the diagnosis and treatment service behavior, it is difficult to select and identify the reasonable process, and it is difficult to obtain the data of the real diagnosis and treatment process in the early research stage. In view of the above problems, this thesis mainly studies the process differentiation, similarity search and process model generation. The research contents and innovations are as follows: (1)Aiming at the problem of monitoring fragmentation of basic medical insurance, this thesis proposes a method based on difference detection between behavioral treatment processes. The treatment process difference detection method based on behavior treats a process model as a sequence of activities, differences in any pair process models can be automatically identified. This thesis first splits the given two process models into small features, which refer to the most common workflow patterns: sequence, split, merge, and loop. Second, it finds out matching activities in the two process models and identifies difference patterns related to matching activities. Third, it identifies behavioral differences between patterns containing matching activities by comparing event logs. Last, it returns all the identified differences between two given process models. For the process whose differences between the reasonable treatment process are many, can be identified as missing or excessive medical services, or even fraud. (2)There is a big gap between the traditional clinical pathway and actual treatment process. Futhermore, it is difficult to select reasonable treatment process from massive cases. To solve this issue, this thesis proposes a medthod to compute similarity between processes based on their structures. This thesis searches through massive treatment processes of the same disease with a classical process, using the structure-based searching method, which considers process models as graphs. Before computing similarity between processes, the method identifies structural differences based on flexible feature matching and returns a list of differences. Then, for each difference, it computes the similarity of two related patterns. Last, it computs an overall similarity between two process models based on similarities between patterns. Using this method, reasonable treatment processes can be selected, which are similar to a given standard process. Later on, if a new treatment process needs to be checked, its similarities are computed with the reasonable processes. The similarity method in this thesis is the first one that computes similarity and returns a list of differences in the business process management area. (3)It is difficult to get real-life treatment processes. Therefore, this thesis proposes a method to compose synthetic process models with properties of real-life processes. This thesis first analyzes the feature set of the treatment process models, which mainly include label feature (label word frequency, word and word co-occurrence probability), structure (probability of sequence, branch, etc.) types (probability of arc types and activities etc.) Then, model fragments are generated according to the characteristics of the analysis, including labels, structures, etc. The resulting fragments are continuously inserted into an empty model according to the probability of the previous analysis, until the size of the process model reaches the desired requirement. Finally, an expected number of synthetic treatment process models are generated. (4)Real-life and synthetic process models are used to evaluate the methods proposed in this thesis. For process difference detection, the experiment shows differeces are found by comparing patterns within 0.013 seconds for a pair of process models on average. For process similarity search, the experiment shows a search among thousands of models is applied within 3.1 seconds and the accuracy of similarity search is over 80%. For process model generation, 10.3 thousands of models are generated within 2 minutes and 42 seconds. All elements in real-life models appear in synthetic models, and these models are similar with regard to their properties. In summary, these methods can be applied to the timely and efficient monitoring of various medical services. Key Words: Difference detection between treatment processes; similarity search among treatment processes; synthesis process model; standardize the process of treatment; process monitoring for medical insurance services

作者相关
主题相关
看过该书的人还在看哪些书