Estimation of Overall Cyclosporine Exposure Using Machine Learning

Background: Cyclosporine (CsA), an immunosuppressant widely used in solid-organ transplantation, requires precise therapeutic drug monitoring to balance its efficacy and toxicity. The interdose area under the concentration–time curve (AUC 0–12 h ) is considered to be a superior metric of drug exposure compared with single concentration measurements but is, nevertheless, resource-intensive. Machine learning (ML) offers a novel approach for AUC prediction by leveraging patient-specific data without relying on traditional pharmacokinetic assumptions. This study intended to develop and evaluate XGBoost ML models for predicting CsA AUC 0–12 h using either two or three blood concentrations and to compare their performance against maximum a posteriori Bayesian estimation (MAP-BE) based on population pharmacokinetic models. Methods: Using data from 2009 patients and 6360 dose-adjustment requests on the Immunosuppressant Bayesian Dose Adjustment website (https://abis.chu-limoges.fr/), supervised ML models were trained with predictors including CsA concentrations at predefined time points (C0, C1, and C3), dose, age, and sampling time deviations. External validation was performed using rich pharmacokinetic profiles of kidney, heart, lung, and bone marrow transplant recipients. Results: The three-sample XGBoost model achieved high accuracy in kidney transplant recipients (root mean square error [RMSE] <3%, RMSE<8.2%), closely matching the MAP-BE performance (rMPE <3%, RMSE <8.7%). The two-sample ML model demonstrated lower precision and higher variability but remained applicable in constrained sampling scenarios. The performance was reduced in heart and lung recipients for both ML and MAP-BE, reflecting the limited representation of these populations in the data set. Conclusions: ML-based AUC prediction is a promising alternative to MAP-BE, particularly for kidney transplantation. Future studies should focus on expanding datasets, incorporating additional transplant types, and refining ML models for broader applicability.