基于BP神经网络、随机森林和决策树建立早期慢性乙型病毒性肝炎肝硬化无创诊断模型

Establishment and comparison of non-invasive diagnostic models for early chronic viral hepatitis B cirrhosis based on BP neural network,random forest and decision tree

目的探讨BP神经网络模型、随机森林模型和决策树CHAID算法模型对早期慢性乙型病毒性肝炎肝硬化的诊断价值。方法收集2018年1月至2019年8月在该院住院且行肝组织活检的106例慢性乙型肝炎病毒(HBV)感染者的临床资料,包括人口学、中医四诊资料、生化指标、肝脏硬度值(LSM)及肝组织病理学结果等参数,采用SPSS18.0软件进行非参数检验和χ^(2)检验,经单因素分析,筛选出对早期肝硬化有统计学差异的指标作为自变量,以有无肝硬化作为因变量,分别建立BP神经网络、随机森林和决策树CHAID算法模型,通过计算3个模型的正确率、错误率、混淆矩阵、灵敏度、特异度、约登指数、阳性预测值、阴性预测值、受试者工作特征(ROC)曲线下面积(AUC)值等来评价和比较3种模型的优劣。结果成功建立了早期慢性乙型病毒性肝炎肝硬化无创诊断模型,在训练组样本和测试组样本中,随机森林算法模型预测正确率(84%,100%)、灵敏度(0.813,1.000)、约登指数(0.701,1.000),ROC AUC(0.896,1.000)均高于BP神经网络算法模型和决策树CHAID算法模型。结论基于随机森林算法的诊断模型建立可以准确判断早期慢性乙型肝炎肝硬化,其预测能力优于BP神经网络模型算法模型和决策树CHAID算法模型,具有良好的临床应用价值。

Objective To explore the diagnostic value of BP neural network model,random forest model and decision tree CHAID algorithm model for early stage chronic viral hepatitis type B(HBV)cirrhosis.Methods The clinical data of 106 chronic HBV-infected patients who were hospitalized in the hospital and had liver tissue biopsy from January 2018 to August 2019 were collected,including parameters such as demographic and TCM four-diagnostic data,biochemical indicators,liver stiffness(LSM)values and liver histopathological findings.Non-parametric tests and chi-square tests were performed by using SPSS18.0 software.After univariate analysis,the indicators that were statistically significant for early cirrhosis were screened as independent variables,and the presence or absence of cirrhosis was used as the dependent variable.BP neural network,random forest and decision tree CHAID algorithm diagnostic models were established respectively.The advantages and disadvantages of the three models were evaluated and compared by calculating the correct rate,error rate,confusion matrix,sensitivity,specificity,Youden index,positive predictive value,negative predictive value and area value receiver operating characteristic(ROC)curve(AUC)of the three models.Results A non-invasive diagnostic model for early stage chronic viral hepatitis type B cirrhosis was successfully established.The correct prediction rate(84%,100%),sensitivity(0.813,1.000),Youden index(0.701,1.000)and AUC(0.896,1.000)of the random forest model were higher than those of the BP neural network algorithm model and the decision tree CHAID algorithm model in both the training group samples and the test group sample.Conclusion The random forest diagnostic model can accurately determine early chronic HBV cirrhosis.Its predictive ability is better than that of the BP neural network algorithm model and the decision tree CHAID algorithm model,and has good clinical application value.