Spatial prediction of human brucellosis (HB) using a GIS-based adaptive neuro-fuzzy inference system (ANFIS).

OBJECTIVE

This study pursues three main objectives: 1) exploring the spatial distribution patterns of human brucellosis (HB); 2) identifying parameters affecting the disease spread; and 3) modeling and predicting the spatial distribution of HB cases in 2012-2016 and 2017-2018, respectively, in rural districts of Mazandaran province, Iran.

METHODS

We collected data on the disease incidence, demography, ecology, climate, topography, and vegetation. Using the Global Moran's I statistic, we measured spatial autocorrelation between log (number of HB cases). We applied the Getis-Ord G statistic to identify areas with high and low risk of the disease. To investigate the relationships between the factors affecting the incidence of HB as input variables together and the factors with the log (number of HB cases) as an output variable, we used the statistical linear regression model and the Pearson correlation coefficient. Then, we implemented a GIS-based adaptive neuro-fuzzy inference system (ANFIS) with two subtractive clustering and fuzzy c-means (FCM) clustering methods to model and predict the spatial distribution of HB.

RESULTS

Global Moran's I spatial autocorrelation analysis indicated that the type of HB distribution is clustered in all years except 2014 and 2017, which are random. According to the Getis-Ord G analysis, the location of the hot spots varied during 2012-2018. In 2012 and 2013, most of the hot spots were seen in the west of the province. While in 2018, they were mostly concentrated in the eastern regions of the province. The linear regression model indicated that the parameters affecting the incidence of HB are independent of each other and can explain only 25.3% of the total changes in the log (number of HB cases). The results of the Pearson correlation coefficient showed that there were positive relationships between vegetation, log (population), and the number of sheep and cattle (p-value < 0.05). The above-mentioned factors had the strongest positive correlation with the log (number of HB cases) (p-value < 0.01). These results may be due to the fact that vegetation regions are suitable for livestock grazing, attracting large crowds of people. Therefore, this will increase HB cases. We compared the results of subtractive clustering and FCM clustering methods by evaluation criteria (e.g., linear correlation coefficient (LCC) and mean absolute error (MAE)) in two phases of development and assessment of the ANFIS model. In the assessment phase, we predicted the spatial distribution of log (number of HB cases) in 2017 and 2018 by subtractive clustering (R = 0.699, LCC or R = 0.692, MAE = 0.509, MSE = 0.455) and by FCM clustering (R = 0.704, LCC or R = 0.697, MAE = 0.512, MSE = 0.448) that showed FCM clustering outperformed the subtractive clustering.

CONCLUSION

The findings may have important implications for public health. The emergence of the hot spots in the east of the province can be a warning to the health system. Health authorities can use the findings of this study to predict the spread of HB and perform HB prevention programs. They can also investigate the factors affecting the prevalence of the disease, identify high-risk areas, and ultimately allocate resources to high-risk regions.