Leveraging Machine Learning Techniques for Variable Selection in Modelling Malaria Transmission

Abstract

Disease modelling is no exception to the widespread acceptance of artificial intelligence across many fields, particularly in cases where a group of highly correlated predictors is linked to the disease's outcome. In order to avoid misleading regression coefficients and inflated standard errors in modelling, multicollinearity must still be absent. With data consisting of predictors that are inherently associated, this study focuses on variable selection in modelling the spread of malaria. Thirteen predictors were analysed, including topography, livestock indices, environmental, and control measure variables. The study addressed multicollinearity in an effort to increase the model's predictive capacity by utilizing machine learning components of artificial intelligence. In particular, regularized algorithms like ridge, elastic net and least absolute and shrinkage selection operator (LASSO) were taken into consideration. A Poisson-based random forest was employed as a comparison tool. There was a small difference between the three regularization method variants under comparison based on the mean square error and Rsquare performance measurements. LASSO outperformed the other two techniques as evidenced by the lowest mean square error value (LASSO = 2.017042, ridge = 2.022117, elastic net = 2.023353). Though the number of important predictors chosen for malaria transmission was precisely the same as that of LASSO, the mean square error of the ensemble Poisson-based random forest (MSE = 0.006167) was much lower than that of the regularisation techniques. The random forest value was twice as high as the regularization algorithm value in terms of R-square. Artificial intelligence is a crucial tool to solve the issues posed by big data as it continues to evolve, especially in the area of disease modelling.