Mycophenolic Acid Exposure Prediction Using Machine Learning.

Area of Science:

• Pharmacology and Toxicology
• Biostatistics and Machine Learning
• Transplant Medicine

Background:

• Therapeutic drug monitoring of mycophenolic acid (MPA) by area under the curve (AUC) is crucial for organ transplant recipients.
• Machine learning (ML) offers potential for more efficient AUC estimation.

Purpose of the Study:

• To develop and evaluate extreme gradient boosting (Xgboost) ML models for estimating MPA AUC0-12h in organ transplant patients.
• To compare the performance of ML models against traditional Bayesian estimation methods.

Main Methods:

• Utilized 12,877 MPA AUC requests from 6,884 patients, developing Xgboost models using two or three concentration measurements.
• Models incorporated concentration differences, sampling time deviations, absorption peak presence, and covariates (dose, transplant type, etc.).
• Evaluated models on test sets and four independent full pharmacokinetic datasets, comparing with MAP Bayesian estimation.

Main Results:

• Xgboost models achieved accurate MPA AUC0-12h estimation with relative bias <5% and relative RMSE <20% on test datasets.
• ML models demonstrated superior performance compared to MAP Bayesian estimation in independent full-PK datasets.
• The models effectively utilized limited concentration data and patient covariates for precise AUC estimation.

Conclusions:

• Xgboost ML models provide accurate and reliable estimation of MPA AUC0-12h, suitable for routine clinical use.
• These models can enhance exposure estimation and facilitate dose adjustment in immunosuppressive therapy.
• The developed ML approach shows promise for integration into clinical decision support systems.