4CPS-208 Machine learning-driven early prediction of voriconazole plasma levels: enhancing precision dosing and patient safety in antifungal therapy

<h3>Background and Importance</h3> Voriconazole, a second-generation triazole with a broad-spectrum of antifungal activity, is essential for treating invasive fungal infections, but its non-linear pharmacokinetics, significant variability between patients, and narrow therapeutic window complicate dosing. Achieving the correct drug concentration is crucial: subtherapeutic levels can result in treatment failure, while elevated levels increase the risk of toxicity. Traditional dosing approaches often lack the precision needed from the start of treatment, and many hospitals lack the resources to routinely perform therapeutic drug monitoring. Machine learning (ML) offers a promising solution to individualise dosing early, improving outcomes and safety. <h3>Aim and Objectives</h3> To develop a machine learning model capable of classifying patients into subtherapeutic, therapeutic, or toxic voriconazole plasma levels early in therapy, enabling clinicians to individualise dosing strategies and minimise the risks of toxicity and subtherapeutic levels. <h3>Material and Methods</h3> A retrospective, single-centre study was conducted between May 2021 and June 2024 at a tertiary hospital. Demographic, clinical, and pharmacokinetic data, including liver and kidney function, serum albumin, comorbidities, and voriconazole plasma concentration (Architect Abbot), were collected. In this study, we propose the use of the extreme gradient boosting (XGB) algorithm for early prediction of voriconazole plasma levels. The performance of XGB was compared to other machine learning methods; K-nearest neighbours (KNN), support vector machines (SVM), Bayesian linear discriminant analysis (BLDA), Gaussian naive Bayes (GNB), and decision trees (DT), demonstrating the superiority of the XGB model. The data were split, 70% used for model training and 30% for validation, applying 5-fold cross-validation to prevent overfitting. <h3>Results</h3> The study included 128 voriconazole-treated patients. XGB achieved 97% accuracy, outperforming KNN (89%) and SVM (86%). BLDA, GNB, and DT showed lower accuracy (81%, 80%, and 84%, respectively). XGB also showed superior results in other metrics, including AUC (0.96), Matthews correlation coefficient (MCC, 85.85%), and Kappa index (86.14%), highlighting its strong performance in classifying subtherapeutic, therapeutic, and toxic levels. <h3>Conclusion and Relevance</h3> This ML-based approach allows early individualisation of voriconazole dosing, offering a precise alternative where traditional methods fall short. By improving dosing accuracy from the start, especially in hospitals without routine therapeutic drug monitoring, this model has the potential to significantly enhance patient safety and treatment efficacy. <h3>References and/or Acknowledgements</h3> <h3>Conflict of Interest</h3> No conflict of interest