Background and Objective <p>Voriconazole is a broad-spectrum antifungal agent whose efficacy and toxicity are closely related to plasma concentrations, which are highly variable between individuals. Therapeutic drug monitoring (TDM) helps optimize its use but is not always available. In this context, machine learning may help predict subtherapeutic or supratherapeutic levels before TDM results are obtained.</p> Methods <p>This was a single-center retrospective study conducted between May 2021 and June 2024 in a tertiary hospital in northern Spain. Adult patients treated with voriconazole for at least 3 days and with a steady-state plasma level measurement were included. Clinical, laboratory, and treatment-related variables were collected. Supervised machine learning models (random forest, support vector machines (SVM), XGBoost, etc.) were trained to classify plasma levels as subtherapeutic, therapeutic, or supratherapeutic.</p> Results <p>A total of 147 patients were included (65% male; median age 65 years). Therapeutic concentrations were found in 71% of patients, supratherapeutic in 15%, and subtherapeutic in 14%. Significant differences were observed on the basis of route of administration, dosage form, age, liver function, and certain comorbidities. Aspartate aminotransferase (AST), glomerular filtration rate, and administration route were the most relevant predictors in the models. Random forest achieved the best performance (area under the curve (AUC) 0.675), though still below the threshold for clinical applicability.</p> Conclusions <p>Although machine learning models identified relevant predictors of voriconazole exposure, their predictive accuracy was limited and insufficient to replace therapeutic drug monitoring. TDM remains essential for individualized and safe dosing. Integrating pharmacogenetic data and hybrid models combining TDM and computational tools may improve predictive performance and clinical applicability.</p>