Ruixiang Liu,1,* Pan Ma,1,* Dongxin Chen,2,* Mengchen Yu,2 Linli Xie,1 Linlin Zhao,2 Yifan Huang,3 Shenglan Shang,2 Yongchuan Chen1 

1Department of Pharmacy, the First Affiliated Hospital of Army Medical University, Chongqing, People’s Republic of China; 2Department of Clinical Pharmacy, General Hospital of Central Theater Command, Wuhan, Hubei Province, People’s Republic of China; 3Medical Big Data and Artificial Intelligence Center, the First Affiliated Hospital of Army Medical University, Chongqing, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Shenglan Shang, Department of Clinical Pharmacy, General Hospital of Central Theater Command, Wuhan, Hubei Province, People’s Republic of China, Email s_shang0818@163.com Yongchuan Chen, Department of Pharmacy, the First Affiliated Hospital of Army Medical University, Chongqing, People’s Republic of China, Email zwmcyc@tmmu.edu.cn

Purpose: Voriconazole (VCZ) is a first-line treatment for invasive fungal disease, characterized by a narrow therapeutic window and significant inter-individual variability. It is primarily metabolized by the liver, the function of which declines with age. Pathological and physiological changes in elderly patients contribute to increased fluctuations in VCZ plasma concentrations. Thus, it is crucial to develop a model that accurately predicts the VCZ plasma concentrations in elderly patients.
Patients and Methods: This retrospective study incorporated 31 features, including pharmacokinetic parameters derived from a population pharmacokinetic (PPK) model. Feature selection for machine learning (ML) models was performed using Recursive Feature Elimination with Cross-Validation (RFECV). Multiple algorithms were selected and combined into an ML ensemble model, which was interpreted using Shapley Additive exPlanations (SHAP).
Results: The predictive performance of ML models was significantly improved by incorporating pharmacokinetic parameters. The ensemble model consisting of XGBoost, random forest (RF), and CatBoost (1:1:8) achieved the highest R2 (0.828) and was selected as the final ML model. Feature selection reduced the number of features from 31 to 9 without compromising predictive performance. The R2, mean absolute error (MAE), and mean squared error (MSE) of the external validation dataset were 0.633, 1.094, and 2.286, respectively.
Conclusion: Our study is the first to incorporate pharmacokinetic parameters into ML models to predict VCZ plasma concentrations in elderly patients. The model was optimized using feature selection and may serve as a reference for individualized VCZ dosing in clinical practice, thereby enhancing the efficacy and safety of VCZ treatment in elderly patients.

Keywords: voriconazole, elderly patients, machine learning, population pharmacokinetics, precision medicine