[Identification of Factors Associated with Supratherapeutic Voriconazole Concentrations Using Decision Tree Analysis]

Voriconazole (VRCZ) is a first-line agent for treating invasive fungal infections, but its pharmacokinetics vary widely, necessitating therapeutic drug monitoring (TDM) to maintain trough concentrations within the therapeutic range (1.0-4.0 µg/mL). C-reactive protein (CRP) is an inflammatory marker reportedly associated with VRCZ concentrations. However, its precise impact in machine learning models remains unclear because of missing values in previous studies. Thus, this study aimed to identify factors associated with supratherapeutic VRCZ concentrations using a classification and regression tree (CART) model incorporating CRP in Japanese patients. We retrospectively analyzed 121 patients who received VRCZ (2-4 mg/kg/dose) and underwent TDM. The patients were classified into a therapeutic range group (1.0-4.0 µg/mL; n=51) and a supratherapeutic group (>4.0 µg/mL; n=70). We excluded outpatients to ensure strict sampling timing, patients whose maintenance doses were changed before the first TDM, and patients with missing laboratory data immediately prior to administration. The CART analysis identified dose (threshold: 2.7 mg/kg/dose) as the primary predictor, followed by CRP (threshold: 5.0 mg/dL) and age (threshold: 74 years). The model demonstrated moderate performance in predicting the therapeutic range group, with an area under the curve of 0.770 [95% confidence interval (CI): 0.689-0.851], sensitivity of 78.4%, and specificity of 71.4%. Dose, CRP, and age were significant factors associated with supratherapeutic VRCZ concentrations. Our findings suggest that a decision-tree-based approach incorporating inflammatory status and age could provide valuable clinical insights for optimizing initial VRCZ dosing.