An Interpretable PK-Informed Hybrid Model for Voriconazole Exposure Prediction: Roles of CYP2C19 Genotype and Inflammation

Yehui Zhou,1,2,* Yuting Yun,1,2,* Shiting Chen,1,2,* Xinyi Liu,2,3,* Hao Liu,4 Jiaxin Ren,1,2 Mengmeng Lu,1,2 Jing Ling,5 Xuping Yang,5 Zhou Zhou,2 James Afriyie Osei,1,2 Respina Soheili,1,2 Jianjun Zou,2 Nan Hu5 1School of Basic Medicine and Clinical Pharmacy, China Pharmaceutical University, Nanjing, People’s Republic of China; 2Department of Pharmacy, Nanjing First Hospital, Nanjing Medical University, Nanjing, People’s Republic of China; 3Department of Clinical Pharmacology, Nanjing First Hospital, Nanjing Medical University, Nanjing, People’s Republic of China; 4State Key Laboratory of Natural Medicines, Key Laboratory of Drug Metabolism, China Pharmaceutical University, Nanjing, People’s Republic of China; 5Department of Pharmacy, The Third Affiliated Hospital of Soochow University, Changzhou, People’s Republic of China*These authors contributed equally to this workCorrespondence: Jianjun Zou, Department of Pharmacy, Nanjing First Hospital, Nanjing Medical University, No. 68 Changle Road, Nanjing, 210006, People’s Republic of China, Email [email protected] Nan Hu, Department of Pharmacy, The Third Affiliated Hospital of Soochow University, Changzhou, 213000, People’s Republic of China, Email [email protected]: Voriconazole (VCZ) exhibits nonlinear pharmacokinetics, a narrow therapeutic window, and substantial interindividual variability. Inaccurate dosing may lead to underexposure or overexposure, causing treatment failure or toxicity. Existing population pharmacokinetic (PPK)-machine learning (ML) models either lack mechanistic interpretability or inadequately characterize VCZ exposure. Therefore, we propose a hybrid model embedding ML within a PPK framework to associate clinical covariates with VCZ exposure.Patients and Methods: A total of 489 inpatients receiving VCZ at the Third Affiliated Hospital of Soochow University between March 2020 and May 2024 were included. We identified candidate predictors of CL/F using dual-feature selection with Boruta and LASSO. Overlapping features were used to train four ML algorithms to estimate CL/F. The predicted CL/F values were incorporated into a steady-state PPK equation to back-calculate VCZ concentrations, followed by quadratic calibration to reduce bias. Causal mediation analysis assessed pathways from key covariates to VCZ concentration via CL/F, and Shapley Additive exPlanations (SHAP) values were used to quantify feature contributions.Results: Under the PK-informed hybrid strategy, XGBoost achieved the best concentration prediction (R2 = 0.739, MAE = 0.357, RMSE = 0.526, MAPE = 7.78%), outperforming a direct ML approach treating PK-related variables as inputs (CatBoost: R2 = 0.459). The temporal external validation performance of the hybrid model remained stable (R2 = 0.661, MAE = 0.473, RMSE = 0.651, MAPE = 14.71%). Mediation analysis demonstrated that CRP affected VCZ exposure primarily through CL/F, whereas albumin and age acted as modifiers. A web-based calculator was developed for real-time individualized prediction and assistance with clinical-dose adjustment.Conclusion: The hybrid model improved VCZ concentration prediction versus direct ML modeling while preserving CL/F-centered mechanistic interpretability. It may help guide dose adjustment and reduce clinically relevant misdosing. This framework may be generalizable to other narrow-therapeutic-window drugs.Keywords: voriconazole, population pharmacokinetics, machine learning, hybrid modeling strategy, causal mediation analysis, plasma concentration prediction