AI-driven modeling of chloroform exposure in pools using PBPK-simulated and experimental data: Implications for environmental exposure science

Predicting chemical uptake under dynamic conditions remains a key challenge in exposure science. This study evaluates two long short-term memory (LSTM) network methods for estimating the alveolar chloroform levels of swimmers in indoor pool facilities. The first approach, which combines mechanistic principles and machine learning capabilities, uses 2000 PBPK simulations to train an LSTM. The second relies exclusively on empirical data and includes six real-world datasets for development and a seventh for independent validation. Both approaches successfully captured the patterns of chloroform absorption and elimination through the respiratory and dermal pathways. Training errors ranged from 0.09 to 0.79, while testing errors were as low as 0.39, a result that reflects the model’s predictive accuracy despite limited real-world data. These findings underscore the potential of both hybrid and experimental data-driven strategies to enhance internal dose predictions in swimming pool environments. This modeling framework facilitates more accurate human health risk assessments in water-based recreational areas and could inform safety guidelines for disinfection by-products in indoor environments.