• ML models were developed to predict lung dosimetry of tobacco product constituents • CFD simulations were performed for sample constituents for use as training data. • The SVR algorithm outperformed RF and XGBoost in predicting vapor uptake in rats. • The SVR rat vapor uptake model was validated against experimental values (R 2 = 0.93). Electronic nicotine delivery systems (ENDS) and combusted cigarettes contain chemical constituents that may be hazardous to human health when inhaled into the respiratory tract. Many of these constituents exist as a vapor when inhaled and may be absorbed into the respiratory tract tissues, but a major challenge in tobacco product risk assessment is that the delivered tissue dose of many constituents in ENDS aerosols remains unknown. To address this gap, machine learning (ML) models were developed to predict vapor uptake and flux of high vapor pressure constituents in the respiratory tract. Latin hypercube sampling (LHS) was used to characterize representative high vapor pressure constituents by sampling over possible ranges of physico-chemical and exposure parameters that affect dosimetry. Computational fluid dynamics (CFD) simulations were performed for each sample constituent to determine vapor uptake and flux in the human nose and mouth and rat nose at six flow rates. The rat uptake data were used to train and compare three ML algorithms: Support Vector Regression (SVR), Random Forest (RF) and eXtreme Gradient Boosting (XGBoost). The SVR model outperformed the other models based on R 2 , RMSE, and MAE. The trained SVR rat vapor uptake model was then validated against vapor uptake of specific chemical constituents from experimental studies (R 2 = 0.93). Thus, the SVR model provides reasonable predictions of vapor uptake and flux for high vapor pressure compounds. These values can be used to derive potential exposures and health risks for the vast number of constituents with unknown exposures following tobacco product use.
Jacketti et al. (2026) studied this question.