Estimation of Overall Cyclosporine Exposure Using Machine Learning.

Area of Science:

• Pharmacology
• Computational Biology
• Transplant Medicine

Background:

• Cyclosporine (CsA) monitoring is crucial for transplant success.
• Interdose area under the concentration-time curve (AUC0-12 h) is a key exposure metric.
• Traditional AUC monitoring is resource-intensive.

Purpose of the Study:

• Develop and evaluate XGBoost machine learning (ML) models for CsA AUC0-12 h prediction.
• Compare ML model performance against maximum a posteriori Bayesian estimation (MAP-BE).
• Assess prediction accuracy using two or three blood concentrations.

Main Methods:

• Trained supervised ML models using patient data (2009 patients, 6360 requests).
• Included CsA concentrations (C0, C1, C3), dose, age, and sampling time as predictors.
• Validated models externally using pharmacokinetic profiles from various transplant recipients.

Main Results:

• Three-sample XGBoost model showed high accuracy in kidney transplant recipients, comparable to MAP-BE.
• Two-sample ML model offered lower precision but was useful in limited sampling situations.
• Performance decreased for heart and lung recipients due to dataset limitations.

Conclusions:

• ML-based AUC prediction is a viable alternative to MAP-BE, especially for kidney transplants.
• Further research should expand datasets and refine ML models for broader use.
• Incorporating diverse transplant types will enhance ML model generalizability.