Estimation of ganciclovir exposure by machine learning

Abstract

11

Estimation of ganciclovir exposure by machine learning

Jean Woillard, Hamza Sayadi, Yeleen Fromage and Selim Arraki Zava

Inserm U1248, Univ Limoges, Chu Limoges

Background: Valganciclovir, an oral prodrug of ganciclovir (GCV), is prescribed to prevent cytomegalovirus infection following transplantation. Dosing adjustments are often based solely on creatinine clearance to achieve a target GCV AUC_0–24 h of 40–50 mg.h/L. However, this approach can result in significant overexposure or underexposure to the drug, potentially compromising efficacy or increasing toxicity. This study aimed to develop and validate machine learning (ML) algorithms capable of accurately estimating GCV AUC, thereby improving dosing precision.

Methods: We simulated 5000 patients for each dosing regimen using two published population pharmacokinetic models (Lalagkas et al., Veniza et al.). These simulated patients were split into training (75%) and testing (25%) datasets. To further evaluate generalizability, an additional validation dataset of 200 patients per regimen was created using two distinct models (Caldés et al., Chen et al.). We developed three ML algorithm configurations using creatinine clearance in combination with either two or three drug concentrations sampled between T0 and T12h and three concentrations restricted between T0 and T6h. The performance of these ML configurations was assessed in both the testing and validation datasets and compared to maximum a posteriori Bayesian estimation (MAP-BE) applied to the Lalagkas et al. and Veniza et al. models within the validation datasets.

Results: Among the ML algorithms evaluated, XGBoost consistently demonstrated the lowest root mean square error (RMSE) during a 10-fold cross-validation, indicating superior predictive accuracy. Models incorporating three blood samples yielded the most precise GCV AUC predictions. In the testing dataset, these models exhibited a relative bias ranging from −0.02% to 1.5% and a relative RMSE between 2.6% and 8.5%. In the validation dataset, the models achieved a relative bias of 1.5% to 5.8% and 8.9% to 16.5%, with a relative RMSE of 8.5% to 9.6% and 10.7% to 19.7% for the Caldés et al. and Chen et al. models, respectively. Notably, the ML algorithm predictions of AUC were significantly more accurate compared to those obtained through the MAP-BE method.

Conclusions: The XGBoost machine learning models provided highly accurate estimates of GCV AUC from as few as two or three blood samples in combination with creatinine clearance. This approach represents a robust limited sampling strategy that can optimize therapeutic drug monitoring, potentially enhancing the clinical management of patients undergoing valganciclovir therapy by reducing the risks associated with drug overexposure and underexposure.