Predictive Map of Spatio-Temporal Distribution of Leptospirosis Using Geographical Weighted Regression and Multilayer Perceptron Neural Network Methods

Detection of pathogenic factors, identify the spatial accumulation of disease cases and finding its distribution pattern are of urgent need in the field of public health and disease management and control. Leptospirosis is a zoonotic disease which occurs worldwide but is most common in tropical and subtropical areas with high rainfall. Wet and mild weather conditions in the northern provinces of Iran have put these areas at high risk for Leptospirosis incidence. The main objectives of this study are to investigate the annual pattern of Leptospirosis distribution, identify the spatial and spatio-temporal clusters of the disease and generate the annual predictive map of spatio-temporal distribution of Leptospirosis at the district level in the Northern provinces of Iran. In this study, Leptospirosis incidences, census data and topographical and climate factors have been used to generate the annual predictive map of spatio-temporal distribution of Leptospirosis. The Leptospirosis incidences are continuously recorded by the Center for Disease Control and Prevention of Ministry of Health of Iran. The population census count estimates for period 2009-2014 were obtained from the Statistical Center of Iran. Topographical data were used to generate the altitude, slope and aspect maps. Climate data such as average temperature, average humidity, annual rainfall and number of freezing days were used to model other affecting parameters. Global clustering methods including Moran’s I and general G indices were applied to investigate the existence of spatial autocorrelation between cases of Leptospirosis and also analyze the annual spatial distribution of the existing patterns. Results of both Moran’s I and general G indices indicated meaningful persistent spatial autocorrelation between Leptospirosis cases and highly clustered distribution of Leptospirosis. Additionally, presence of spatial clusters of Leptospirosis and detection of high risk areas of disease were investigated using the local Moran’s I and local G indices. The results of the local Moran’s I and local G indices identified significant spatial clusters of Leptospirosis cases located in central, north-eastern and western parts of Guilan, Mazandaran and Golestan provinces, respectively. Geographically weighted regression (GWR) and multilayer perceptron neural network (MLP) models were used to generate the annual predictive map of spatio-temporal distribution of Leptospirosis and modelling the relation between the distribution of Leptospirosis cases with topographical and climate factors. Performance of GWR and MLP models were compared using Kappa coefficient, RMSE, MAPE and R2 measures. The evaluation results showed that the MLP model was able to predict the incidence rate of Leptospirosis in 2014 for each district with acceptable accuracy. MLP was able to model the relationship between Leptospirosis incidence and factors better than GWR. Additionally, results of both GWR and MLP models showed that average humidity and annual rainfall are most important affecting factors on Leptospirosis incidence in the Northern provinces of Iran, respectively. Such predictive maps can be used to provide essential guidelines for planning of effective control strategies and identification of high risk areas of Leptospirosis which should receive more preventive measures from policy makers and healthcare authorities.