Model Ensembling and Machine Learning Approaches to Predict the First Dose of Amoxicillin in Intensive Care.

Area of Science:

• Pharmacology
• Pharmacometrics
• Computational Biology

Background:

• A priori model-informed precision dosing (MIPD) aims to optimize initial drug dosages using patient covariates, avoiding the need for concentration measurements.
• Population pharmacokinetic (PopPK) models are crucial for understanding drug behavior in diverse patient groups, particularly in critical care settings.

Purpose of the Study:

• To develop and evaluate novel model ensembling and machine learning (ML) strategies for predicting optimal amoxicillin first doses in intensive care unit (ICU) patients.
• To compare the performance of these advanced MIPD methods against traditional dosing strategies and single PopPK models.

Main Methods:

• Simulated a virtual patient population using data from four published amoxicillin PopPK models.
• Developed and implemented weighted model ensembling (WME), classification tree (CT)-informed ensembling, regression tree (RT)-informed ensembling, and factor analysis of mixed data (FAMD).
• Trained ML algorithms (SVM, k-NN, Random Forest, XGBoost) to predict doses achieving target concentrations based on patient covariates and dosing schemes.

Main Results:

• Most MIPD methods, including ensembling and ML, demonstrated superior performance over standard dosing and single-model PopPK approaches in simulated data.
• Ensembling methods achieved 24%-40% correct predictions, while ML methods reached 36%-39%, outperforming single-model approaches (17%-32%).
• ML-based ensembling showed promise in increasing target attainment in both simulated and clinical data, reducing the need for explicit model selection.

Conclusions:

• Advanced MIPD strategies, particularly ML-based ensembling, offer a significant improvement for amoxicillin dosing in intensive care settings.
• These methods facilitate faster achievement of therapeutic drug concentrations, enhancing patient outcomes.
• The primary limitation for broader application is the scarcity of high-quality clinical datasets required for robust model training and validation.