BP神经网络模型用于气象因素对脑出血死亡影响的初步研究

摘要

目的 探讨BP神经网络预测模型在分析气象因素与脑出血死亡率关系中的应用.方法 根据BP神经网络的特性,利用MATLAB 7.0软件的神经网络工具箱对2007-2009年哈尔滨市气象数据建立脑出血死亡率的BP神经网络预报模型,并与传统的多元线性回归模型进行比较.结果 利用多元线性回归结果显示脑出血死亡率与最高气温、最小相对湿度呈负相关,与平均相对湿度、日照时数呈正相关.脑出血死亡率的非线性相关系数(RNL)为0.7854,平均绝对误差百分比( MAPE)为0.21,均方误差(MSE)为0 22,平均绝对识差(MAE)为0.19,预测准确度(P)为81.31％,平均误差率为0.19.BP神经网络模型的拟合结果显示,脑出血死亡率的RNL为0 7967,MAPE 为0.19,MSE为0.21,MAE为0.18,P为82.53％,平均误差率为0.17.结论 应用BP神经网络预测模型对2010年哈尔滨市脑出血死亡率进行预报,通过与多元线性回归模型预报结果进行比较,表明该模型具有更高的预报准确度.%Objective Using the Back Propagation (BP) Neural Network Model to discover the relationship between meteorological factors and mortality of intracerebral hemorrhage,to provide evidence for developing an intracerebral hemorrhage prevention and control program,in Harbin.Methods Based on the characteristics of BP neural network,a neural network Toolbox of MATLAB 7.0 software was used to build Meteorological data of 2007-2009 with intracerebral hemorrhage mortality to predict the effect of BP neural network model,and to compare with the traditional multivariate linear regression model. Results Datas from the multivariate linear regrcssion indicated that the cerebral hemorrhage death mortality had a negative correlation with maximum temperatureand minimum humidity while having a positive correlation with the average relative humidity and the hours of sunshine.The linear correlation coefficient of intracerebral hemorrhage mortality was 0.7854,with mean absolute percentage (MAPE) as 0.21,mean square error (MSE) as 0.22,mean absolute error(MAE) as 0.19.The accuracy of forecasting was 81.31％ with an average error rate as 0.19.The Fitting results of BP neural network model showed that non-linear correlation coefficient of intracerebral hemorrhage mortality was 0.7967,with MAPE as 0.19,MSE as 0.21,MAE as 0.18.The forecasting accuracy was 82.53％ with the average error rate as 0.17.Conclusion The BP neural network model showed a higher forecasting accuracy when compared to the multiple linear regression model on intraccrebral hemorrhage mortality,using the data of 2010' s.