Providing a model for Cutaneous Leishmaniasis risk mapping using GIS and neural network algorithm

Cutaneous Leishmaniasis (CL) is a vector-borne disease, endemic of the Middle East. The spread of CL is highly associated with the socio-ecological interactions of vectors, hosts and environmental conditions. CL is the most frequent vector-borne disease in Iran and especially in the north-eastern province, Golestan, which has long been known as one of the most important endemic areas for CL dispersion. Therefore, Golestan province was selected as the study area of this research. The main objectives of the study are to analyze annual spatial distribution of CL, investigate the relations between environmental/climate factors and incidence rate of CL and also provide a model to predict CL distribution at rural district level in Golestan province. Data: CL incidences, census data, environmental and climate factors have been used in this study to provide a model and produce a map to predict the CL distribution. The CL incidences are continuously recorded by the Center for Disease Control and Prevention (CDC) of Golestan province. The population and census data for 2013-2015 period were obtained from Iranian Statistical Center. Environmental and climate data such as vegetation, average humidity, average temperature, precipitation, number of rainy days, number of freezing days, maximum wind speed and evaporation rate were used as parameters affecting the model. The statistical and geo-statistical analyses were used to investigate the relation between environmental/climate factors and CL incidence rate, and to investigate the existence of spatial autocorrelation between CL cases, respectively. Additionally, Multilayer perceptron (MLP) neural network was used to model the relation between the distribution of CL incidences with environmental/climate factors, and also to generate the risk maps of CL. MLP is a type of neural network which consists of multiple layers of neurons or processing elements connected in a feed forward fashion. It encompasses three types of layers: input, hidden, and output. It has a unidirectional flow of information. Generally, information flow starts from input layer, goes through hidden layer, and then to output layer, which provides the response of the network to the input stimuli. In this type of network, there are generally three distinct types of neurons in layers. The input layer contains some neurons as the input variables. The hidden neurons, which are contained in one or more hidden layers, process and encode information within the network. The hidden layer receives, processes, and passes the input data to the output layer. Number of hidden layers and number of neurons within each layer affect the accuracy and functionality of the network. The output layer contains target output vector. In this study, effective parameters along with CL incidence rate of 2013-2014 were fed to the MLP as training data. The trained MLP was used afterward to generate the risk map of 2015 and test accuracy of the model. In order to determine the optimal parameters of the MLP, the grid-search and cross-validation techniques were used on 25% of the training dataset in the training phase. The performance of MLP was investigated using the root mean square error (RMSE), mean absolute percentage error (MAPE) and area under curve (AUC) of receiver operating characteristic (ROC) measures. Sensitivity analysis was also used to determine most effective variables regarding predictive mapping of CL distribution.  Results of global Moran’s I index indicated that there is spatial autocorrelation among CL cases, and also distribution of CL cases in Golestan province in each 3 years is clustered. Moreover, statistical analyses showed that majority of the incidences belonged to rural districts of Gonbad-Kavos and Maraveh-Tappeh. Based on the results of statistical analyses (including Pearson correlation and Spearman rank correlation), positive correlations were observed between the CL incidence rate and average temperature, maximum wind speed and evaporation. In addition, negative correlation was found between the CL incidence rate and average humidity, precipitation, number of rainy days, number of freezing days and vegetation. According to the results of evaluation criteria including RMSE, MAPE and AUC, the trained MLP model was able to generate risk maps of CL in 2013-2015 for each rural district with acceptable accuracy. Additionally, results of sensitivity analysis indicate that vegetation and average humidity are the most influencing variables in the incidence of CL and in predictive mapping of CL distribution in Golestan province.
Conclusion and Future works:
In this study, the global Moran’s I index indicated the presence of spatial autocorrelation among CL cases, and clustered distribution of disease in the study area. The statistical analyses showed that environmental and climate factors greatly affect the spatial distribution of CL. The MLP method, used to generate CL distribution risk maps, was able to generate the study area risk maps with acceptable accuracy. Results highlight the potential high risk areas requiring special plans and resources for monitoring and control of the disease. As a future work, we suggest that the effects of other environmental and socio-economic parameters should be evaluated to improve the accuracy of the model. It is also recommended that other methods such as regression and other neural network techniques be used to generate CL risk maps.