Predicting busulfan exposure in patients undergoing hematopoietic stem cell transplantation using machine learning...

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Machine Learning in Medicine
• Drug Monitoring

Background:

• Busulfan (BU) is a critical chemotherapeutic agent.
• Accurate prediction of busulfan area under the curve at steady state (AUCss) is essential for optimizing patient outcomes.
• Traditional population pharmacokinetic (pop PK) models may have limitations in individualizing BU therapy.

Purpose of the Study:

• To develop and compare machine learning (ML) models for predicting busulfan (BU) area under the curve at steady state (AUCss).
• To identify the optimal ML algorithm for accurate BU AUCss prediction.
• To evaluate the performance of ML models against a conventional population pharmacokinetic (pop PK) model.

Main Methods:

• A retrospective study of 79 adult patients receiving intravenous BU was conducted.
• The dataset was split into training (80%) and testing (20%) groups.
• Nine ML algorithms and one pop PK model were developed and validated to predict BU AUCss.

Main Results:

• All developed ML models demonstrated superior performance compared to the pop PK model in both model fitting and predictive accuracy.
• The ML models achieved higher R-squared values and lower Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) than the pop PK model.
• Support Vector Regression (SVR) and Gradient Boosted Regression Trees (GBRT) algorithms yielded the best predictive ability for BU AUCss, with R² of 0.953.

Conclusions:

• Machine learning models, particularly those using SVR and GBRT, offer superior predictive accuracy for busulfan AUCss compared to traditional pop PK models.
• These advanced ML models can facilitate individualized busulfan dosing strategies.
• The findings support the potential clinical application of ML for optimizing busulfan therapy and improving patient care.