人工神经网络在耐甲氧西林葡萄球菌感染患儿万古霉素个体化给药中的应用  

Application of artificial neural network in vancomycin individualized administration in children with methicillin- resistant Staphylococcus infections

作　　者：汪洋[1] 张华年[1] 徐华[1] 陈渝军[1] 刘茂昌[1] 

机构地区：[1]武汉市儿童医院,湖北武汉430016

出　　处：《中国新药与临床杂志》2015年第2期137-143,共7页Chinese Journal of New Drugs and Clinical Remedies

基　　金：武汉市科技攻关计划(2013060602010258)

摘　　要：目的建立人工神经网络模型用于估算耐甲氧西林葡萄球菌(MRS)感染患儿万古霉素稳态血药浓度,以指导个体化给药。方法收集100例MRS感染患儿静脉泵注万古霉素后的180例次稳态血药浓度数据和临床资料。将所有血药浓度数据和病例资料随机分成两组,训练组(n=150)采用遗传算法配合动量法训练后建立人工神经网络模型,另外建立多元线性回归模型;测试组(n=30)用建立的人工神经网络预测测试组患儿的血药浓度,通过计算平均预测误差(MPE)、权重残差(WRES)、平均绝对预测误差(MAE)、平均预测误差平方(MSE)和均方根预测误差(RMSE)来验证模型。结果人工神经网络MPE(0.33±1.86)mg·L-1,WRES(14.83±14.55)%,MAE(1.38±1.26)mg·L-1,MSE(3.45±5.32)(mg·L-1)2,RMSE 1.86 mg·L-1;人工神经网络模型有83%的血药浓度数据绝对预测误差<3.0 mg·L-1,而多元线性回归模型仅有53%。人工神经网络预测的准确度及精密度均优于多元线性回归模型。结论本研究建立的人工神经网络预测性能较好,可用于预测MRS感染患儿万古霉素稳态血药浓度以指导个体化给药。AIM To establish an artificial neural network （ANN） for predicting vancomycin concen- trations in children with methicillin-resistant Staphylococcus （MRS） infections and optimize the individual use of vancomycin. METHODS Totally 180 cases of steady state plasma concentration data were collected from 100 MRS- infected children with intravenous infusion of vancomycin in this study. All the concentrations and data were divided into two groups randomly. In the training group （n =150）, ANN was established after the network parameters were trained by using momentum method combined with genetic algorithm, and multiple linear regression （MLR） model was established. In the testing group （n = 30）, the concentrations of the testing group patients were predicted by ANN established, and the mean predicted error （MPE）, weighted residuals （WRES）, mean absolute prediction error （MAE）, mean squared prediction error （MSE）, root mean squared prediction error （RMSE） were calculated to assess the ANN model. RESULTS The assessed results of ANN were MPE （0.33±1.86） mg·L-1, WRES （14.83 ±14.55） %, MAE （1.38 ± 1.26） mg·L-1, MSE （3.45 ± 5.32） （mg·L-1） 2, RMSE 1.86 mg·L-1, respectively. There were 83% of ANN test data sets whose absolute prediction errors were less than 3.0 mg·L-1, while the percent of MLR model was 53%. The accuracy and precision of ANN were superior to those of MLR with stepwise method. CONCLUSION The established ANN in this study is good enough to predict vancomycin concentration and optimize its individualization of dosage regimen in children with MRS infections.

关 键 词：万古霉素 神经网络(计算机) 甲氧西林抗药性 葡萄球菌属 儿童 个体化给药 

分 类 号：R969.1[医药卫生—药理学]