Comparative modelling approaches of personal exposure to particle-associated PAH

Abstract

Several models for simulation of personal exposure (PE) to particle-associated polycyclic aromatic hydrocarbons (PAH) have been developed and tested. The modeling approaches include linear regression models (Model 1), time activity weighted models (Models 2 and 3), a hybrid model (Model 4), a univariate linear model (Model 5), and machine learning technique models (Model 6 and 7). The hybrid model (Model 4),whichutilizes microenvironment data derived from time-activity diaries (TAD) with the implementation of add-on variables to account for external factors that might affect PE, proved to be the best regression model (R2 for B(a)P ) 0.346, p < 0.01; N )68). This model was compared with results from two machine learning techniques, namely decision trees (Model 6) and neural networks (Model 7), which represent an innovative approach to PE modeling. The neural network model was promising in giving higher correlation coefficient results for all PAH (R2 for B(a)P ) 0.567, p < 0.01; N ) 68) and good performance with the smaller test data set (R2 for B(a)P ) 0.640, p < 0.01; N ) 23). Decision tree accuracies (Model 6) which assess how precisely the algorithm can determine the correct classification of a PE concentration range indicate good performance, but this is not comparable to the other models through R2 values. Using neural networks (Model 7) showed significant improvements over the performance of hybrid Model 4 and the univariate general linear Model 5 for test samples (not used in developing the models). The worst performance was given by linear regression Models 1 to 3 based solely on home and workplace concentrations and time-activity data.