Integrating machine learning with aggregate risk models to assess acute myeloid leukemia from chemical mixture exposure

Traditional aggregate risk models rely on the availability of toxicological parameters for each substance in the mixture and assume a linear relationship between the substances. However, these models are limited by incomplete toxicological data and the potential for non-linear (antagonistic or synergistic) interactions among substances. To address these limitations, this work presents a framework for characterizing and interpreting toxicological data to support health risk prediction. Toxicological parameters were obtained from the United States Environmental Protection Agency's Integrated Risk Information System database, and exposure data from an artificially generated dataset based on the properties of exposure data published by Heck et al. The latter dataset is highly imbalanced, with a ratio of negative to positive individuals of 0.0024 of total samples, thus posing modelling challenges. Specifically, we apply Principal Component Analysis (PCA) to infer relationships among exposure substances, identify clusters in the principal component space to estimate missing toxicological parameters, and use the resulting PCA loadings to cross-check linear aggregation weights learned via gradient descent. We then extend the linear risk aggregation model to include pairwise interaction terms to capture non-linear dependencies and compare its interpretability and predictive performance against three tree-based models: Random Forest (RF), Balanced Random Forest (BRF), and eXtreme Gradient Boosting (XGBoost). Our results show that PCA reveals groupings of high-risk pollutants that align with both the established literature and the gradient-learned linear aggregation weights. Interaction-based extensions to the linear model produced unstable and uninterpretable results, likely due to extreme data imbalance and the limitations of using pairwise exposure products to represent interactions. Among tree-based methods, BRF achieved the best performance on imbalanced data, and using a minority oversampling technique to balance the data was insufficient to achieve strong predictive performance on the extremely imbalanced datasets typically encountered in toxicological risk prediction applications.