Advanced explainable machine learning approaches provide insight into complex contaminant interactions.

Traditional environmental health research often focuses on the toxicity of exposure to a single chemical, but in the real world, people are exposed to multiple pollutants simultaneously, which may interact in complex ways to amplify or mitigate toxic effects.

Traditional models that assume additive effects such as additive concentrations or independent actions can be misleading in these scenarios. Advanced statistical and machine learning methods have been employed to address this issue, but they are often inadequate due to the complexity, the large number of interacting contaminants, and the inability to extract the absolute effect of each contaminant.

To address this issue, a research group led by Professor Kuk Cho at Busan National University introduced Feature Localized Intercept Transformed-Shapley Additive Explanations (FLIT-SHAP) as a solution to these challenges. This tool is unique in that it analyzes the impact of specific contaminants within a mixture, unlike traditional methods that use a broader approach.

This detail-oriented technique is particularly useful for understanding the OP of airborne particulate matter (PM), which is known to cause a variety of health problems. Hazardous Materials Journal.

“The OPs of PM refer to the ability of these particles to generate reactive oxygen species, which cause oxidative stress and inflammation in the body, leading to various health issues such as respiratory and cardiovascular diseases,” explains Professor Cho. Using FLIT-SHAP, the scientists look at the OPs of these particles in mixtures of pollutants. The lab-curated data was used to train robust machine learning models such as eXtreme Gradient Boosting (XGBoost), and the FLIT-SHAP explanatory tool was used to uncover key insights.

“In controlled laboratory environments, we found that 55-63% of interactions were synergistic (working together to increase toxicity) and 25-42% were antagonistic (counteracting each other). However, in real-world scenarios, antagonistic effects were more common, and overall toxicity was lower than predicted by traditional models,” Prof Cho added.

One surprising finding was the significant impact of a group of chemicals called quinones, which contributes more to toxicity than previously thought. This finding calls into question current approaches to regulating pollutants and suggests that more focus should be placed on these chemicals. FLIT-SHAP not only predicted toxicity with excellent accuracy (R² = 0.99), but also provided detailed insight into how different contaminants interact with each other. This makes it a powerful tool for conducting realistic risk assessments and improving our understanding of environmental health risks.

The researchers' rigorous testing revealed how contaminants interact at different concentrations and revealed that traditional additive models often over- or underestimate the toxicity of mixtures. Through FLIT-SHAP, the researchers identified important nuances in the contribution of contaminants in different environments, highlighting the need for more refined risk assessments.

“While this study focuses on specific contaminants and toxicity endpoints, further research on broader scenarios is essential. FLIT-SHAP represents a major leap forward, providing an accurate tool for assessing the toxicity of chemical mixtures. This advancement is expected to lead to better regulatory decisions to effectively protect human health.”