Comparison of the Use of Multilayer Perceptron Neural Network with Weighted Regression in Determining the Geographical Pattern of the Brucellosis Disease Incidence in Mazandaran Province, Iran (2009-2018)

Background and purpose: Time series models based on machine learning, including artificial neural network, perform better than classical methods. This study was performed to compare the use of a multi-layered perceptron neural network with weighted regression in determining the geographical pattern of brucellosis in Mazandaran province, Iran (2009-2018) on 3,732 patients. Materials and methods: The study type is ecological and existing data, which is registered. Multilayer perceptron neural network was used to model and predict disease occurrence, and the results were compared with weighted regression. The root mean square error (RMSE) and mean absolute percentage error (MAPE) were used to compare the models. All analyses were performed using SPSS software (version 26) and Microsoft Excel 2016. Results: The age-standardized rate of disease incidence was 13.2 per 100,000. The highest incidence rate (17.2) was in 2012 and the lowest (10.6) in 2014. Galugah city (35.0) and Qaemshahr and Fereydunkenar cities (3.0) had the highest and lowest incidence rates. The independent variables studied included male gender, rural residence, age over 55 years, contact with dairy products, being a homemaker, rancher and associated with livestock jobs, contact with livestock, and livestock vaccination on the standard incidence of brucellosis in both artificial neural network and weighted regression (other than variable over 55 years) models were effective. Conclusion: The predictive power of the model in the multi-layered perceptron neural network was higher than the weighted regression.