Predicting the Frequency of Human Brucellosis using Climatic Indices by Three Data Mining Techniques of Radial Basis Function, Multilayer Perceptron and Nearest Neighbor: A Comparative Study

Identification of statistical models has a great impact on early and accurate detection of outbreaks of infectious diseases and timely warning in health surveillance. This study evaluated and compared the performance of the three data mining techniques in time series prediction of brucellosis.

In this time series, the data of the human brucellosis cases and climatology parameters of Hamadan, west of Iran, were analyzed on a monthly basis from 2004 (March/April) to 2017 (February/March). The data were split into two subsets of train (80%) and test (20%). Three techniques, i.e. radial basis function (RBF) and multilayer perceptron (MLP) artificial neural network methods as well as K Nearest neighbor (KNN), were used in both subsets. The root mean square errors (RMSE), mean absolute errors (MAE), mean absolute relative errors (MARE), determination coefficient (R2) and intra-class correlation coefficient (ICC) were used for performance comparison.

Results indicated that RMSE (23.79), MAE (20.65) and MARE (0.25) for MLP were smaller compared to the values of the other two models. The ICC (0.75) and R2 (0.61) values were also better for this model. Thus, the MLP model outperformed the other models in predicting the used data. The most important climatology variable was temperature.

MLP can be effectively applied to diagnose the behavior of brucellosis over time. Further research is necessary to detect the most suitable method for predicting the trend of this disease