基于随机森林回归模型的登革热风险评估究

目的 基于随机森林回归模型构建小空间尺度的登革热风险评估工具,为登革热防控提供依据。方法 以2012年1月至2014年9月登革热病例及相关因素数据为训练集,分别构建登革热流行频率、持续时间及强度风险指标的随机森林回归模型,以2014年10月至2015年12月登革热病例及相关因素数据为验证集,并对构建的模型进行评估。结果 频率、持续时间、强度指标与发病数指标的相关系数均>0.7。依据训练集构建的登革热流行频率、持续时间和强度风险指标的随机森林回归模型变量解释度分别为96.72%、91.98%和90.1%,提示模型拟合度较好;交叉验证法可见各模型均方误差分别0.001 9、1.424 6和1.881 1,均处于较低水平;比较随机森林回归、支持向量回归、广义线性模型和广义相加模型的准确性,随机森林回归和支持向量机等机器学习模型均方误差远低于广义线性模型和广义相加模型。结论 以登革热频率、持续时间及强度指标为结局变量,气象、环境及社会经济特征为预测变量构建的随机森林回归模型准确性较好,可作为登革热风险评估工具,为登革热防控工作服务。

Risk assessment of dengue fever based on random forest model

Objective To construct a small spatial scale dengue risk assessment tool based on the random forest model,so as to provide scientific basis for the prevention and control of dengue fever. Methods Data of dengue case and related factors from February 2012 to September 2014 were used as the training set and random forest regression （RFR） models were constructed separately for frequency, duration and intensity of dengue fever. Data of dengue cases and related factors from October 2014 to March 2015 were used to as the testing set to verify the accuracy of the models. Results The correlation coefficients between incidence and frequency, duration, intensity of dengue fever were all higher than 0.7. Based on the training set, the pseudo R-squareds in the models of frequency, duration, and intensity were 96.72%, 91.98%, and 90.1%; the cross-validated mean square errors （MSEs） of the models were 0.001 9, 1.424 6, and 1.881 1, respectively. By comparing the accuracy of RFR, support vector regression （SVR）, generalized linear model （GLM） and generalized additive model （GAM）, the MSEs of RFR and SVR were much lower than those of GLM and GAM. Conclusion The RFR models constructed using the frequency, duration and intensity of dengue fever as outcome variables and the meteorological, environmental and socioeconomic characteristics as predictors have better accuracy and can be used as a risk assessment tool for preventing and control of the outbreak of dengue fever.