Background Voriconazole (VCZ) is a first-line treatment drug for invasive fungal disease, with narrow therapeutic window and significant inter-individual variability. VCZ is primarily metabolized by liver, which declines with age. For elderly patients, physiological and pathological factors lead to more pronounced fluctuations in VCZ plasma concentrations. Thus, it is crucial to establish a model that accurately predicts VCZ plasma concentration in elderly patients. Methods A retrospective study was performed incorporating 32 variables, including population pharmacokinetics (PPK) parameters derived from the PPK model. Recursive Feature Elimination with Cross-Validation was used for feature selection. Multiple algorithms were selected and combined into an ensemble model, and the model was interpreted by Shapley Additive exPlanations. Results The predictive performance of machine learning was greatly improved after inclusion of PPK parameters. The composition of XGBoost, RF, and CatBoost (1:1:8) with the highest R2 (0.828) was determined as the final ensemble model. Feature selection greatly simplified the model from 31 variables to 9 variables without compromising its performance. The R2, mean square error, mean absolute error, and accuracy (± 30%) of external validation were 0.633, 1.094, 2.286, and 71.05%, respectively. Conclusions Our study is the first to include PPK parameters as new factors for machine learning modeling to predict VCZ plasma concentrations in elderly patients. The model underwent optimization through feature selection. Our model provides a reference for individualized dosing of VCZ in clinical practice, enhancing the efficacy and safety of VCZ treatment in elderly patients. Keywords: voriconazole, elderly patients, machine learning, population pharmacokinetics, precision medicine