Blending Bayesian Networks with Mechanistic Models to Improve Exposure Risk Predictions for Environmental Contaminants

Jacqueline MacDonald Gibson, Ph.D.
Twisdale Family Civil, Construction, and Environmental Engineering Department Head and Professor
North Carolina State University

Abstract

Rationale

For decades, environmental exposure modeling has relied on mechanistic models grounded in physics, chemistry, and biology to predict the fate and transport of pollutants. A classic example is the Streeter-Phelps model, developed in the 1920s to evaluate dissolved oxygen depletion from wastewater discharges. These models continue to play a vital role in regulatory decision-making, including setting air and water quality standards, evaluating pesticide applications, and prioritizing households for lead (Pb) remediation. However, they often require extensive parameterization and may be limited in their predictive power, particularly in complex or data-rich contexts. Bayesian networks offer under-explored opportunities to improve exposure risk predictions. Yet the relative strengths and weaknesses of mechanistic, Bayesian network, and hybrid approaches remain poorly characterized in real-world settings.

Approach

We compared the performance of mechanistic, machine-learned Bayesian network, and hybrid modeling approaches for predicting household-level exposure risks to environmental contaminants in two case studies. In the first, we evaluated predictions of PFAS concentrations in private well water using measurements from 1,205 North Carolina households. In the second, we modeled the risk of children’s exposure to lead (Pb) in drinking water, indoor dust, and outdoor soil using data from over 100,000 children with blood lead measurements. Across both cases, model performance was evaluated based on classification accuracy, false negative rate, and predictive power for identifying households with exposures exceeding health-based guidelines.

Results and Discussion

For PFAS exposure, Bayesian network models outperformed a mechanistic transport model across all accuracy metrics and were significantly more powerful in avoiding false negatives. A hybrid model that incorporated mechanistic predictions into the Bayesian network achieved the highest accuracy and the lowest false negative rate. Preliminary results for the lead exposure analysis suggest similar advantages for integrated modeling, with full findings to be presented at the conference.

Management and Policy Implications

As George Box famously noted, “All models are wrong, but some are useful.” Our findings suggest that models may become even more useful when mechanistic frameworks are integrated with machine learning approaches trained on observational data. This integration holds promise for improving the accuracy of risk assessments, enhancing early identification of at-risk populations, and informing regulatory and policy decisions regarding environmental health hazards.

About the Presenter

Dr. Jacqueline MacDonald Gibson serves as the Twisdale Family Head of the Department of Civil, Construction, and Environmental Engineering at North Carolina State University, where she leads a dynamic engineering department that is engaged in a wide range of big problems that affect everyday lives—from designing safer skyscrapers, airports, and roadways to ensuring access to safe, affordable drinking water and clean air. With a foundation of 13 years in high-impact public policy roles before transitioning to academia, Dr. Gibson’s research integrates advanced engineering methods to identify innovative solutions for mitigating environmental risks and shaping evidence-based public policy.

Her distinguished public policy career includes serving as Associate Director of the U.S. National Research Council’s Water Science and Technology Board and acting as a key liaison to the White House Office of Science and Technology Policy during her tenure with The RAND Corporation. Her interdisciplinary research addresses critical global challenges, ranging from optimizing water infrastructure to enhance public health in the United States to developing strategic frameworks for prioritizing environmental policy initiatives in the United Arab Emirates.

A recognized leader in the field, Dr. Gibson is the President of the Society for Risk Analysis and serves as an Associate Editor for Environmental Science & Technology, one of the premier journals in environmental engineering and science. She holds dual PhDs in Engineering and Public Policy and in Civil and Environmental Engineering from Carnegie Mellon University, an MS in Environmental Science in Civil Engineering from the University of Illinois at Urbana-Champaign, and a BA in Mathematics from Bryn Mawr College. Dr. Gibson’s career exemplifies the power of interdisciplinary expertise to address pressing environmental challenges and influence transformative policy decisions worldwide.