基于不同对照抽样方法下的全球高致病性禽流感H5N1空间风险分析

目的  基于全球高致病性禽流感H5N1(highly pathogenic avian influenza virus H5N1, HPAIV H5N1)疫情数据，探讨比较四种对照抽样对全球HPAIV H5N1空间风险分析的影响。方法  疫情数据来源于官方疫情监测报告。构建以国家为单位的完全随机抽样、缓冲区抽样、基于人口密度的概率抽样和最大熵(MaxEnt)抽样四种抽样，按1∶4比例抽样作为对照数据。收集整理距铁路、公路和候鸟迁徙路径的最短距离、土地类型、全球数字高程模型、婴儿死亡率6个因素，采用Logistic空间自回归模型进行建模，基于受试者工作特征曲线下面积(the area under the curve, AUC)、灵敏度、特异度等评价指标，比较不同抽样的影响并预测风险。结果  四种抽样所得的AUC值在0.896~0.971间，预测能力均好，而MaxEnt抽样预测能力最优。从预测风险来看，随机抽样和缓冲区抽样预测结果存在偏差，概率抽样的预测结果存在低估，MaxEnt抽样空间风险预测结果最优。结论  随机抽样下的全球HPAIV H5N1疫情空间风险建模结果最差，MaxEnt抽样的建模结果最好，空间风险预测地区更为准确，可作为全球禽流感空间流行病学研究中合理对照选择的参考，同时建议在今后类似的研究中，须重视对照抽样方法的选择。

Spatial risk analysis of global highly pathogenic avian influenza H5N1 virus based on different methods of choosing controls

  Objective  To explore and compare the impacts of four sampling methods on the spatial risk analysis of global highly pathogenic avian influenza virus H5N1 (HPAIV H5N1) based on the global databases of HPAIV H5N1.  Methods  H5N1 outbreak data were collected from the official surveillance reports. The control data were obtained through four sampling methods using the ratio of 1∶4 for case to control, completely random sampling based on the country, buffer zone sampling, probability sampling based on population density and probability sampling based on the results of MaxEnt. Six influencing factors were collected, including the shortest distance from the epidemic point to the railway, highway and wild bird migration routes, land use and land cover data, altitude, and infant mortality rate. The Logistic regression model with spatial autocorrelation term was applied to analyze and compare the impacts of four sampling methods on the predicted risk through the area under the curve (AUC), sensitivity, specificity and other evaluation indicators.  Results  For the four sampling methods, their AUC values were between 0.896-0.971, showing their prediction ability were good, and the MaxEnt-based sampling method had the best predictive ability. From the perspective of predicting risk, the results of random sampling and buffer zone sampling were biased, and the result of probability sampling was underestimated, while the results for MaxEnt sampling was best.  Conclusion  The results of spatial risk modeling for global HPAIV H5N1 based on random sampling was poorest, but MaxEnt sampling had the best modelling result, whose predicted risk regions were more accurate. Hence it can provide a rational reference to select controls for the spatial epidemiologic researches of global avian influenza. It is suggested to pay attention to the selection of control sampling methods in future similar researches.