Comparative modeling approaches for personal exposure to particle-associated PAH

Area of Science:

• Environmental Health
• Exposure Science
• Computational Toxicology

Background:

• Personal exposure (PE) simulation is crucial for understanding health risks associated with particle-associated polycyclic aromatic hydrocarbons (PAH).
• Existing models vary in complexity and accuracy, necessitating evaluation of novel approaches.

Purpose of the Study:

• To compare the performance of various models for simulating personal exposure to particle-associated PAH.
• To evaluate the efficacy of machine learning techniques against traditional regression and hybrid models.

Main Methods:

• Development and testing of multiple simulation models: linear regression, time-activity weighted, hybrid, univariate linear, decision trees, and neural networks.
• Utilized microenvironment data from time-activity diaries and external factors for the hybrid model.
• Employed R-squared values and correlation coefficients to assess model performance.

Main Results:

• The hybrid model (Model 4) showed good regression performance (R² for B(a)P = 0.346).
• Neural networks (Model 7) demonstrated superior performance, yielding higher correlation coefficients for all PAH (R² for B(a)P = 0.567) and strong results on test data (R² for B(a)P = 0.640).
• Decision trees (Model 6) showed good classification accuracy but were not directly comparable using R² values. Linear regression models performed worst.

Conclusions:

• Neural network models represent a significant advancement in simulating personal exposure to particle-associated PAH.
• Machine learning, particularly neural networks, offers improved accuracy over traditional and hybrid models for PE assessment.
• Further research into machine learning applications for exposure science is warranted.