Association Between Air Pollution and Child Brain Development: A Machine Learning Based Environmental Modeling Study
Description
This study examines the relationship between air pollution exposure and child brain development using machine learning–based environmental modeling. Air pollution is increasingly recognized as a determinant of neurodevelopmental outcomes, particularly during early life when the brain undergoes rapid structural and functional changes. The project integrates simulated multi-source pollution data with advanced analytical techniques to characterize complex, nonlinear exposure–response relationships. A synthetic dataset of 100 exposure scenarios was developed to represent realistic variations in pollutant levels originating from industrial operations, vehicular traffic, dust storms, and wildfires. Pollutants included NOx, CO, CO₂, dust PM₁₀, and wildfire PM₂.₅. Emission levels were generated within typical environmental ranges. A Total Exposure Index was calculated as a weighted sum of pollutants to represent cumulative exposure intensity. The Predicted Impact (%) variable was simulated using a nonlinear function incorporating pollutant interactions and controlled random noise, ensuring reproducibility through fixed random seeds. Each dataset entry was labeled with one of several machine learning techniques (Random Forest, XGBoost, SVM, Neural Networks, Gradient Boosting, Linear Regression) to illustrate methodological diversity. Analyses were performed using Python and standard open-source libraries in a reproducible Jupyter Notebook environment. Exploratory analysis included pairwise scatter plots, correlation heatmaps, and boxplots to visualize relationships and compare model behaviors. A Random Forest model with cross-validation was applied to assess predictive performance and feature importance. Results indicated that the Total Exposure Index was the dominant predictor, followed by particulate matter and CO. Correlation analysis revealed a strong positive relationship between total exposure and predicted impact, emphasizing the significance of overall pollutant burden in shaping modeled neurodevelopmental outcomes. Although synthetic, this dataset demonstrates the utility of machine learning approaches in environmental health research. It provides a structured, transparent, and reproducible framework that can be adapted to real-world monitoring data. The methodology supports method development, sensitivity analysis, and policy-relevant modeling, aligning with current trends in data-driven environmental epidemiology. By highlighting key exposure drivers and interactions, the study contributes to a growing evidence base informing public health protection and regulatory decision-making in the context of vulnerable populations, particularly children. The approach can also be extended to diverse geographical regions or pollutant mixtures. This enhances its value as a flexible educational and research resource.
Files
Steps to reproduce
The following steps outline the procedures used to generate, process, and analyze the dataset, enabling full reproducibility of the study: Data Generation A synthetic dataset of 100 exposure scenarios was created to represent realistic spatial–temporal variations in air pollutant levels originating from industrial operations, vehicular traffic, dust storms, and wildfires. Pollutants included NOx, CO, CO₂, dust PM₁₀, and wildfire PM₂.₅. Concentration ranges were based on typical industrial and urban emission levels to maintain environmental plausibility. A Total Exposure Index was calculated as a weighted sum of pollutants to represent cumulative exposure intensity, with weighting factors reflecting relative source contributions. Outcome Simulation A Predicted Impact (%) variable was generated using a nonlinear mathematical function that accounted for pollutant interactions, CO₂ effects, and controlled random noise. This produced realistic yet reproducible exposure–impact relationships. Fixed random seeds were applied to ensure deterministic results across repeated runs. Machine Learning Labeling Each dataset entry was randomly assigned one of several modeling techniques (e.g., Random Forest, XGBoost, SVM, Neural Networks, Gradient Boosting, Linear Regression). This labeling illustrates how multiple analytical approaches can be applied to the same environmental dataset for methodological comparison. Analytical Workflow The dataset underwent structured exploratory analysis to examine pollutant distributions, interrelationships, and the relationship between exposures and predicted impact. Visual assessments included scatter plots, correlation maps, and comparative summaries across different analytical techniques to highlight patterns and variability. Modeling and Validation A structured modeling approach was applied using cross-validation to assess predictive performance and feature importance. Key exposure drivers were ranked through feature importance, and statistical correlations were computed between numeric features and the predicted outcome. The Total Exposure Index consistently emerged as the dominant predictor, highlighting its integrative role in the dataset. Reproduction Instructions To replicate: Generate the dataset using the described pollutant ranges and nonlinear impact formula. Compute the Total Exposure Index and Predicted Impact. Perform exploratory analyses to understand variable relationships. Apply analytical models with cross-validation to evaluate predictive strength. Examine feature rankings and correlations to interpret results. Consistency can be ensured by maintaining the same random seeds, weighting factors, and analytical procedures. The methodology is fully transparent, allowing researchers to adapt and extend it to various contexts or real-world datasets.
Institutions
- University of Calgary