Different modelling approaches of personal exposure to selected volatile organic compounds (VOC)

Abstract

Volatile organic compounds (VOCs) are known to have various health effects on the quality of human life, of which some compounds are known or suspected to be carcinogenic to humans. The most common and accurate way of determining personal exposure (PE) is done by direct measurements, however, these experiments are expensive and time consuming to carry out. Therefore, modelling using data obtained from filled-in questionnaires and time activity diaries is a very good alternative. The aim of this dissertation is to try and improve the current PE models and determine the factors or activities which mostly contribute to particular levels of VOC concentrations. This was done by two main methods, namely machine learning techniques (MLT) and using general linear models (GLM). Five MLT were considered including decision trees, artificial neural networks and k-nearest neighbour. These are made to learn or train on 75% of the dataset and then use this knowledge to predict the outcome to new unseen data, which for this case would be the PE level or concentration. However, a Univariate GLM uses another approach. This requires the data to be normally distributed which is easily done by applying some kind of transformation. When the model is validated on the 25% testing set, for benzene in MATCH, results reach R2 = 0.434 when predicting continuous data, and also reaching 86.46% accuracy when predicting nominal data (classifications). The main focus in this dissertation was set on benzene and 1,3-butadiene due to their carcinogenicity as well as toluene due to its high concentrations, and details on the modelling approaches used and results obtained are discussed in the main text.