Predicting nicotine emissions and plasma nicotine boost in E-cigarette users using machine learning

As e-cigarette use continues to rise globally, the need for reliable and scalable methods to quantify nicotine exposure in e-cigarette users has become more crucial. This study aims to develop supervised machine learning (ML) models that predict nicotine emission and plasma nicotine boost based on puff topography, personal and device characteristics, and e-liquid consumption. This study analyzed 259 measurements from e-cigarette users aged 21-35 years by combining human puff topography data with laboratory toxicant emission tests generated through validated playback procedures. Puffing behavior was recorded during ad libitum vaping sessions and replicated with a puffing robot (playback) to measure nicotine emissions. Plasma nicotine levels were determined via liquid chromatography-mass spectrometry. Six supervised regression models, including ordinary least squares, lasso, random forest, XGBoost, support vector regression, and neural networks, were trained to predict nicotine emissions and plasma nicotine boost from puffing parameters. The models were evaluated using an 80/20 train, test split with bootstrap resampling over 200 iterations, with performance metrics assessed by R², RMSE, and MAE. Among the ML methods evaluated, XGBoost consistently demonstrated the best performance (highest R², minimal RMSE, and MAE) for predicting both nicotine emissions and plasma nicotine boost. For nicotine emission prediction, XGBoost achieved the highest accuracy using either puff number (R² = 0.778 ± 0.153) or liquid consumption (R² = 0.747 ± 0.149) in combination with average puff duration and device characteristics (nicotine concentration and device). Using these emission estimates together with individual characteristics (age, sex, height, and weight), XGBoost-based plasma nicotine boost models showed the best explanatory performances for both puff-based (R² = 0.613 ± 0.174) and liquid-consumption-based (R² = 0.699 ± 0.168) predictors. This study shows that ML methods can effectively estimate nicotine emissions and plasma nicotine boost exposure directly from puffing behavior and device features. These models can support product regulation, assess addiction risk, and facilitate population-level surveillance.Trial registration: ClinicalTrials.gov (NCT05205382, NCT05338801).