A prediction model of occupational manganese exposure based on artificial neural network

Application of two statistical models to reconstruct occupational exposure to manganese (Mn) is discussed. Air monitoring of 635 samples were analyzed by a back-propagation artificial neural network (back-propagation ANN) in comparison with a multiple linear regression (MLR). The stepwise MLR yielded significant results with five selected variables for predicting airborne manganese dioxide (MnO(2)). However, a 6-12-1 back-propagation ANN was superior to the data from MLR. Statistical parameters and non-parametric paired tests indicated that back-propagation ANN represents the more useful and accurate tool. ANN was used to predict missing MnO(2) concentrations in the present study. The median of MnO(2) was 0.445 mg/m(3) (IQR 0.131-1.342). The MnO(2) characteristics of time, distance, and exposure site were defined. Airborne MnO(2) for three previous periods (1978-1988, 1989-1998, and 1998-2007) were 1.228 mg/m(3), 0.664 mg/m(3), and 0.501 mg/m(3), respectively. The medians were 0.350 mg/m(3), 0.281 mg/m(3), and 0.190 mg/m(3) at distances of 5, 10, and 25 m away from the site of exposure. Compared with levels encountered in other studies, mine concentrator sites were more seriously polluted, due to the practices of direct ore processing.