中国媒介生物学及控制杂志

• 论著 • 上一篇    下一篇

广义回归神经网络在肾综合征出血热发病率预测中的应用

吴伟1;郭军巧2;王萍3;周宝森1   

  1. 1中国医科大学流行病学教研室 沈阳110001;2辽宁省疾病预防控制中心;3沈阳市疾病预防控制中心
  • 出版日期:2007-06-20 发布日期:2007-06-20

Application of generalized regression neural network in forecasting incidence of hemorrhagic fever with renal syndrome

WU Wei; GUO Jun-qiao; WANG Ping; ZHOU Bao-sen   

  1. Department of Epidemiology, School of Public Health, China Medical University, Shenyang 110001, China
  • Online:2007-06-20 Published:2007-06-20

摘要: 目的 探讨广义回归神经网络(GRNN)在肾综合征出血热(HFRS)发病率预测上的优势及应用前景。方法 利用1984-2002年沈阳市的气象资料(包括平均气温、相对湿度、降水量和日照)和动物疫情资料(包括鼠密度和鼠带病毒率)共6个指标作为神经网络的输入,将1985-2003年沈阳市HFRS发病率作为神经网络的输出。利用Matlab7.0软件中的神经网络工具箱分别构建HFRS发病率的GRNN预测模型和反馈(BP)神经网络预测模型,对样本进行拟合和预测并对两者的拟合和预测性能进行比较。结果 GRNN的最优光滑因子为0.35;BP神经网络的隐含层数定为6。从拟合效果来看,GRNN和BP神经网络预测模型的平均误差率(MER)分别为25.42%和25.55%;两者的决定系数r2分别为0.9438和0.9729,总的来说,拟合效果比较满意,两者拟合差异不是很明显。从预测效果来看,两者的MER分别为4.90%和15.16%,GRNN的MER远远小于BP神经网络;两者的r2分别为0.9897和0.9516。结论 GRNN充分体现了它在小样本预测中的优势,预测效果优于BP神经网络,对解决HFRS等流行情况影响因素复杂的问题有很好的实用价值。

关键词: 广义回归神经网络, 反馈神经网络, 肾综合征出血热, 预测

Abstract: Objective To study the superiority and application prospect of generalized regression neural network(GRNN) which is used in forecasting the incidence of hemorrhagic fever with renal syndrome(HFRS). Methods Use meteorological data, including average temperature, relative humidity, precipitation and sunshine time, and epidemiologic information of animal diseases, including rodent density and viral carriage of rodents from 1984 to 2002 as the input of neural network. Use the incidence of HFRS from 1985 to 2003 as the output of neural network. Construct the GRNN forecasting model and BP neural network forecasting model respectively with the neural network toolbox of Matlab7.0. Fit and forecast the sample and compare the performance between the two different neural networks. Results The optimize smooth factor of GRNN is 0.35; the hidden layers of BP neural network is 6. From the fitting effect, the MER of GRNN and BP neural network are 25.42% and 25.55% respectively; their r2 are 0.9438 and 0.9729. On the whole, the fitting effect is satisfactory, and the difference of the two neural networks is not very significant. From the forecasting effect, the MER between the two neural networks are 4.90% and 15.16% respectively. The MER of GRNN is less than the MER of BP neural network; their r2 are 0.9897 and 0.9516. Conclusion GRNN is more superior in small sample forecasting than BP neural network, and the forecasting effect is better. GRNN has practical value in solving epidemic problem which has complicated influencing factor such as HFRS.