Evaluation of Voriconazole Population Pharmacokinetics in Japanese Patients and Factors Influencing Prediction Accuracy: A Machine Learning Approach

Abstract

Voriconazole (VRCZ) exhibits highly variable pharmacokinetics. Although CYP2C19 poor metabolizers are common in Japan, routine genetic testing is rarely performed in daily clinical practice. Consequently, existing population pharmacokinetic (Pop-PK) models that lack genetic data require external validation in Japanese clinical settings. This study aimed to identify the most accurate nongenetic Pop-PK model for this population and elucidate the factors influencing prediction errors. We retrospectively analyzed 174 adult inpatients who received VRCZ within the first 14 d of therapy. Eight Pop-PK models were evaluated using prediction error metrics and specific clinical criteria. For the best-performing model, normalized prediction distribution error (NPDE) analysis was conducted. Subsequently, a classification and regression tree (CART) analysis with leave-one-out cross-validation (LOOCV) was performed to identify factors associated with prediction accuracy. Model H, incorporating the albumin–bilirubin (ALBI) score, performed best (relative root mean square error 2.97, criteria 33.91%, median prediction error (MDPE) −29.06%). However, NPDE analysis (mean 0.774, variance 2.394) revealed residual systematic underprediction and substantial unmodeled variability. CART analysis identified C-reactive protein (CRP) as the primary determinant of accuracy. While the best model (age, platelet count, serum creatinine, and CRP) achieved a training accuracy of 69.5%, internal validation via LOOCV yielded an accuracy of 56.9% and a multi-class area under the receiver operating characteristic curve of 0.568. While the ALBI-score–based Model H is the most accurate, predictions remain limited in high-inflammation states due to inherent stochastic variability. Stratifying patients by baseline CRP level is crucial for optimizing the initial VRCZ dosing.