基于数据挖掘技术建立的BP神经网络模型鉴别儿童川崎病与发热性疾病的研究

发布时间：2019-07-11 14:42

【摘要】：目的以临床表现和实验室指标建立诊断川崎病(KD)的BP神经网络模型并考察其诊断性能。方法收集重庆医科大学附属儿童医院(我院)2007年1月至2016年1月电子病历系统中出院诊断为KD的连续病例和待鉴别发热疾病病例,使用R 3.2.3软件中的随机抽样函数分为训练集和测试集。截取病历中一般情况、临床表现和实验室指标的共51项信息,单因素分析后提取有统计学意义的变量,以此分别构建Logistic回归和BP神经网络模型,比较两种模型的诊断性能。结果 905例KD患儿和438例待鉴别发热疾病患儿进入数据模型分析,训练集1 042例,其中KD 700例,待鉴别发热类疾病342例;测试集301例,其中KD 205例,待鉴别发热类疾病96例。单因素分析结果显示差异有统计学意义37项信息。Logistic回归分类模型有16个变量纳入最佳回归方程。BP神经网络输入层、隐含层和输出层分别有37、24和1个节点。Logistic回归分类模型对训练集和测试集的分类正确率为84.1%和82.1%,ROC曲线下面积为0.91和0.89;BP神经网络模型对训练集和测试集的分类正确率为96.4%和86.0%,ROC曲线下面积为0.94和0.92;2个模型的敏感度均很好,BP神经网络模型的特异度优于Logistic回归分类模型。结论本文建立的BP神经网络诊断模型对KD有较好的诊断辅助功能,有待进一步通过临床检验。
[Abstract]:Objective to establish a BP neural network model for the diagnosis of (KD) in Kawasaki disease based on clinical manifestations and laboratory indexes and to investigate its diagnostic performance. Methods from January 2007 to January 2016, the continuous cases of KD diagnosed in the electronic medical record system of affiliated Children's Hospital of Chongqing Medical University (our hospital) and the cases of febrile diseases to be identified were collected. The random sampling function in R 3.2.3 software was used to divide them into training set and test set. A total of 51 items of information about general situation, clinical manifestation and laboratory index were intercepted, and statistically significant variables were extracted after univariate analysis, and Logistic regression and BP neural network models were constructed to compare the diagnostic performance of the two models. Results 905 children with KD and 438 children with febrile diseases to be identified entered the data model analysis. There were 1042 cases of training set, including 700 cases of KD and 342 cases of febrile diseases to be identified, including 301 cases of KD and 96 cases of febrile diseases to be identified. The results of univariate analysis showed that there were 37 items of information with statistical significance. 16 variables in Logistic regression classification model were included in the optimal regression equation. BP neural network input layer, hidden layer and output layer were 37, 24 and 1 node, respectively. The classification accuracy of Logistic regression classification model for training set and test set was 84.1% and 82.1%, and the area under ROC curve was 0.91 and 0.89. The classification accuracy of BP neural network model for training set and test set is 96.4% and 86.0%. The sensitivity of the two models under ROC curve is 0.94and 0.92. the specificity of BP neural network model is better than that of Logistic regression classification model. Conclusion the BP neural network diagnostic model established in this paper has a good diagnostic auxiliary function for KD, and needs to be further tested.
【作者单位】：重庆医科大学医学信息学院;重庆医科大学附属儿童医院心内科;
【基金】：重庆市自然科学基金:cstc2015shmszx0301
【分类号】：R725.4

【相似文献】