Advancing exposure science through artificial intelligence: Neural ordinary differential equations for predicting blood concentrations of volatile organic compounds

Abstract

Volatile organic compounds (VOCs) are a significant concern for human health and environmental safety, requiring accurate models to predict their concentrations in body fluids for effective risk assessments. This study evaluates the application of neural ordinary differential equations (Neural ODEs), a novel artificial intelligence (AI)-based framework, in predicting blood concentrations of two representative VOCs: dibromomethane (DBM) and methylchloroform (MCF). Using data across multiple concentrations, Neural ODEs demonstrated robust predictive performance, achieving lower mean absolute percentage errors (MAPEs) than conventional Physiologically Based Pharmacokinetic (PBPK) models in most scenarios. For DBM, Neural ODEs achieved a MAPE of 6.56 % at 10,000 ppm, outperforming PBPK in low-to-moderate exposure levels. Neural ODEs yielded a MAPE of 25.55 % for MCF at 10,000 ppm, though accuracy declined at lower concentrations, such as 10 ppm. The findings emphasize the adaptability of Neural ODEs for diverse cases while discussing potential advancements such as integrating real-time monitoring tools and hybrid modeling approaches. This study underscores the value of Neural ODEs as a versatile tool for toxicological forecasting, environmental monitoring, and regulatory decision-making.