Development of an explainable machine learning asthma prediction model using serum brominated flame retardants in a national population.

We aimed to explore the association of serum brominated flame retardant (BFR) metabolites and mixture profiles with asthma risk among US adults. Data were sourced from the National Health and Nutrition Examination Survey (NHANES), 1999-2023. Four machine learning methods (light gradient boosting machine, eXtreme gradient boosting [XGBoost], random forest, and neural network) annexed with SHapley Additive exPlanations (SHAP) and one traditional logistic regression were used to develop and validate an explainable asthma prediction model. This study included 9,948 US adults. XGBoost outperformed other models with the highest area under the curve (AUC) at 0.814. Sixteen features-family history, BMI, PBDE47, PBDE28, PBDE154, age, race/ethnicity, smoking, second-hand smoking, sex, education, PIR, marriage, drinking, PBDE153, and PBB153-identified by at least two of applied methods were ultimately entered into machine learning models. According to the SHAP-quantified contribution to asthma risk, five key BFRs in predicting asthma were identified: PBDE47, PBDE28, PBDE154, PBDE153, and PBB153. Our findings indicated that XGBoost model proved most effective in predicting adulthood asthma based on serum BFRs. This machine learning-based model holds substantial promise for the early prevention, risk stratification, and clinical management of asthma.