Improving risk models for avian influenza: the role of intensive poultry farming and flooded land during the 2004 Thailand epidemic

Abstract

Since 1996 when Highly Pathogenic Avian Influenza type H5N1 first emerged in southern China, numerous studies sought risk factors and produced risk maps based on environmental and anthropogenic predictors. However little attention has been paid to the link between the level of intensification of poultry production and the risk of outbreak. This study revised H5N1 risk mapping in Central and Western Thailand during the second wave of the 2004 epidemic. Production structure was quantified using a disaggregation methodology based on the number of poultry per holding. Population densities of extensively- and intensively-raised ducks and chickens were derived both at the sub-district and at the village levels. LandSat images were used to derive another previously neglected potential predictor of HPAI H5N1 risk: the proportion of water in the landscape resulting from floods. We used Monte Carlo simulation of Boosted Regression Trees models of predictor variables to characterize the risk of HPAI H5N1. Maps of mean risk and uncertainty were derived both at the sub-district and the village levels. The overall accuracy of Boosted Regression Trees models was comparable to that of logistic regression approaches. The proportion of area flooded made the highest contribution to predicting the risk of outbreak, followed by the densities of intensively-raised ducks, extensively-raised ducks and human population. Our results showed that as little as 15% of flooded land in villages is sufficient to reach the maximum level of risk associated with this variable. The spatial pattern of predicted risk is similar to previous work: areas at risk are mainly located along the flood plain of the Chao Phraya river and to the south-east of Bangkok. Using high-resolution village-level poultry census data, rather than sub-district data, the spatial accuracy of predictions was enhanced to highlight local variations in risk. Such maps provide useful information to guide intervention.

Figure 1. Distribution of HPAI H5N1 outbreaks and flooded areas.

The distribution of flooded areas (A) and Higly Pathogenic Avian Influenza outbreaks (B) follow a comparable distribution pattern in the Central-Eastern region of Thailand and at local scale in Phitsanulok and Phetchabun provinces (C). The grey zone represents the area excluded from model training due to absence of poultry data.

Figure 2. Villages “bubble” pattern in Nakhon Ratchasima Province.

Villages are identified as point-based administrative units of Thailand. Their area was artificially delimited by intersecting Thiessen polygons with a one kilometre radius circular buffer.

Figure 3. Predicted risk of HPAI outbreak.

Maps of average predicted risk at the village level (A) and at the sub district level (B) for 25 iteration of the Boosted Regression Trees model. (C) and (D) represent the corresponding uncertainty maps (coefficient of variation, log scaled) associated with the risk prediction. The grey zone represents the area where poultry data were unavailable for training models.

Figure 4. Local risk predictions (sub-district vs village).

A comparison of local risk maps highlight the added value of village level prediction compared to the continuous risk surface based on sub-district poultry data. Village level poultry data allows improved targeting of potential intervention measures.

Figure 5. Predictors relative contribution to HPAI H5N1 risk model.

Number of crop cycles (n.crops), fraction of a village neighborhood/sub-district covered with lake water (lake), human population density (Hpop, log10 scale), fraction of a village neighborhood/sub-district covered with river water or floods (floods), residual spatial autocorrelation (RAC), number/density per square kilometer of extensively raised chickens (Ext.Ch), intensively raised chickens (Int.Ch), extensively raised ducks (Ext.Du) and intensively raised ducks (Int.Du). Flooded areas and intensively raised duck shows the highest contribution to model at the village level (A) whereas at the sub-district level (B) the intensively raised ducks density is the main determinant of risk of outbreak. The relative contributions are based on the number of times a variable is selected for a node in the Boosted Regression Trees model weighted by the squared improvement to the model as a result of each node and averaged over all trees. Contributions are scaled so that the sum adds to 100.

Figure 6. Relationship between risk factors and HPAI H5N1 fitted risk function at the village level.

The HPAI H5N1 risk functions of the BRT models are plotted for the number of crop cycles (n.crops; A), fraction of a village neighborhood covered with lake water (lake; B), human population density (Hpop, log10 scale; C), fraction of a village neighborhood covered with river water or floods (floods; D), residual spatial autocorrelation (RAC, E), number of extensively raised chickens (Ext.Ch; F), intensively raised chickens (Int.Ch; G), extensively raised ducks (Ext.Du; H) and intensively raised ducks (Int.Du; I). The grey lines present the predicted line for each of the 25 iterations and the black line is the average prediction.

Figure 7. Relationship between risk factors and HPAI H5N1 fitted risk function at the sub-district level.

The HPAI H5N1 risk functions of the BRT models are plotted for the number of crop cycles (n.crops; A), fraction of a village neighborhood covered with lake water (lake; B), human population density (Hpop, log10 scale; C), fraction of a village neighborhood covered with river water or floods (floods; D), residual spatial autocorrelation (RAC; E), number of extensively raised chickens (Ext.Ch; F), intensively raised chickens (Int.Ch; G), extensively raised ducks (Ext.Du; H) and intensively raised ducks (Int.Du; I). The grey lines present the predicted line for each of the 25 iterations and the black line is the average prediction.