基于贝叶斯网络在肝硬化并发肝性脑病相关因素及分类识别的应用研究

发布时间：2018-04-19 02:39

本文选题：贝叶斯网络 + 肝性脑病　；参考：《山西医科大学》2017年硕士论文

【摘要】：目的:建立肝硬化并发肝性脑病相关因素的贝叶斯网络模型,探索肝性脑病与这些可能因素间的关系,通过贝叶斯网络推理反映一个或多个因素对肝性脑病的作用强度;尝试构建肝性脑病分类识别模型,探讨贝叶斯网络用于肝性脑病识别的分类效果,为临床医生识别肝性脑病提供合理的方法,为肝性脑病的智能识别奠定前期基础。方法:收集2006年1月~2015年12月在山西医科大学第一附属医院消化内科住院治疗并具有完整病历资料的950例肝硬化患者,利用单因素及多因素logistic回归分析筛选出肝硬化并发肝性脑病的相关因素,构建贝叶斯网络拓扑结构。然后在单因素筛选出肝性脑病相关因素的基础上,采用禁忌搜索算法构建贝叶斯网络(GBN),探讨GBN用于肝性脑病识别的分类效能,同时与朴素贝叶斯分类器(NB)、树扩展朴素贝叶斯分类器(TAN)及logistic概率预测模型的识别能力进行比较。结果:1.将与肝性脑病有关系的23个因素进行单因素及多因素logistic回归分析,结果显示最终进入回归模型的因素有7个,电解质紊乱,感染,精神萎靡,肝肾综合征,肝源性糖尿病及凝血酶原时间延长、总胆红素升高均与肝性脑病有关;其中,电解质紊乱、肝肾综合征、感染是肝硬化患者并发肝性脑病的主要危险因素,其风险分别为6.861倍、3.467倍、3.021倍;精神萎靡、肝源性糖尿病、凝血酶原时间延长、总胆红素升高与HE的关系相近,相对危险度在2.1~2.7范围。2.构建的贝叶斯网络ROC曲线下的面积为0.843,网络评价效果较好。各因素通过复杂的关系与肝性脑病建立联系,其中肝肾综合征、电解质紊乱、感染、精神萎靡、总胆红素和凝血酶原时间与肝性脑病有直接关系,肝源性糖尿病通过总胆红素间接与肝性脑病发生关联。网络推理发现,感染、电解质紊乱和肝肾综合征与肝性脑病的关系更密切。3.根据本研究950例肝硬化数据集的分类识别效果评价可看出,禁忌搜索算法构建的贝叶斯网络(GBN)对肝性脑病的整体识别效能高于其他模型(F-measure为0.410,G-measure为0.739);经过1:2抽样,缩小两类间样本数的不平衡现象后,GBN的G-measure值为0.754,仅次于TAN,而GBN的F-measure值依然高于其它模型(F-measure值为0.820)。相对于950例不平衡数据集,GBN的G-measure值增加了2.0%,F-measure值增加了1倍,说明GBN对肝性脑病的整体识别效能有所提高,尤其对肝硬化并发肝性脑病的阳性识别性能提升幅度较大,在约登指数最大的条件下,概率截断点由0.10提高为0.374,该水平在识别HE时更为合理。结论:禁忌搜索算法构建的贝叶斯网络灵敏度特异度高,能反映各节点之间的关系,能够揭示相关因素间的联系及对肝性脑病的作用,贝叶斯网络推理可以根据医生掌握患者信息的先后顺序,对患者并发肝性脑病进行推理,符合临床诊疗序贯过程。尝试用禁忌搜索算法构建的贝叶斯网络(GBN)用于肝硬化并发肝性脑病的识别,其对肝性脑病阳性分类识别能力高于其它模型,对肝性脑病的筛查可能有一定的指导意义,但是用于临床肝硬化并发肝性脑病分类识别效果还需要外部数据的验证。
[Abstract]:Objective: to establish the Bayesian network model related factors of liver cirrhosis complicated with hepatic encephalopathy, to explore the relationship between hepatic encephalopathy and the possible factors, by the Bayesian network inference reflects one or more factors on the strength of hepatic encephalopathy; try to construct the hepatic encephalopathy classification recognition model of Bayesian network for classification of hepatic encephalopathy provides recognition. A reasonable method for clinicians to identify hepatic encephalopathy, will lay the foundation for intelligent recognition of hepatic encephalopathy. Methods: from January 2006 December ~2015 in 950 cases of liver cirrhosis patients in the First Affiliated Hospital of Shanxi Medical University Department of Gastroenterology Hospital treatment and with complete medical records, analysis showed that the related factors of liver cirrhosis complicated with hepatic encephalopathy using univariate and multivariate logistic regression, to construct a Bayesian network topology. Then the single factor screening of hepatic encephalopathy. Based on the factors of constructing the Bias network using tabu search algorithm (GBN), to investigate the classification performance of GBN for hepatic encephalopathy recognition, and at the same time Naive Bayesian classifier (NB), extended Naive Bayesian tree classifier (TAN) model recognition ability to compare prediction and probability of logistic. Results: 23 factors and 1. the relationship between hepatic encephalopathy logistic univariate and multivariate regression analysis, the results show the final factors into the regression model 7, electrolyte disorder, infection, listlessness, hepatorenal syndrome, hepatic diabetes and prolonged prothrombin time, total bilirubin increased associated with hepatic encephalopathy; among them, electrolyte disorder hepatorenal syndrome, infection is a major risk factor in patients with liver cirrhosis complicated with hepatic encephalopathy, the risk was 6.861 times, 3.467 times, 3.021 times; listlessness, hepatogenic diabetes, prothrombin time. Long, total bilirubin increased close relationship with HE, Bias network ROC curve of relative risk in the range of 2.1~2.7.2. under the construction area of 0.843, better network evaluation effect. Various factors through the complicated relationship with hepatic encephalopathy which establish contact, hepatorenal syndrome, electrolyte disorder, infection, listlessness, total bilirubin direct and the relationship between prothrombin time and hepatic encephalopathy, hepatogenic diabetes through total bilirubin encephalopathy and indirect association. Network inference, infection, electrolyte disorder and hepatorenal syndrome, hepatic encephalopathy and more closely.3. according to the classification results in this study, 950 cases of liver cirrhosis of the evaluation data set can be seen, the tabu search algorithm Bias network the construction of the overall recognition performance (GBN) of hepatic encephalopathy is higher than other models (F-measure = 0.410, G-measure = 0.739); after 1:2 sampling, narrow between two classes Imbalance of sample number, GBN G-measure value is 0.754, second only to TAN, GBN and F-measure value is still higher than that of other models (F-measure = 0.820). Compared with 950 cases of imbalanced data sets, GBN G-measure value increased by 2%, F-measure increased by 1 times, indicating that the overall recognition performance of hepatic GBN encephalopathy has increased, especially the positive recognition performance of liver cirrhosis complicated with hepatic encephalopathy improved greatly, the Youden index under the condition of the maximum probability, the cut-off point was improved from 0.10 to 0.374, the level of recognition in HE is more reasonable. Conclusion: the sensitivity of the Bayesian network can avoid constructing search algorithm for high specificity, can reflect the relationship between between the various nodes, can reveal the relationship between related factors of hepatic encephalopathy and the role of Bayesian network inference can be based on the order of the doctor grasp of the information of the patients, for patients with hepatic encephalopathy In accordance with the clinical diagnosis and treatment of sequential reasoning process. Try the search algorithm for Bayesian network constructed by Tabu (GBN) for identification of liver cirrhosis complicated with hepatic encephalopathy, the hepatic encephalopathy positive recognition ability than other models, screening for hepatic encephalopathy may have a certain guiding significance, but for clinical liver cirrhosis complicated with hepatic encephalopathy classification effect need to verify the external data.

【学位授予单位】：山西医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R575.2;R747.9

【参考文献】