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

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

Problem Statement
The paper addresses the challenge of achieving rapid target attainment for amoxicillin in intensive care unit (ICU) patients without relying on prior concentration measurements. While Model-Informed Precision Dosing (MIPD) offers a priori dose recommendations based on patient covariates, the selection of a single Population Pharmacokinetic (PopPK) model often leads to suboptimal predictions due to model uncertainty and variability in patient populations. The study seeks to improve the accuracy of first-dose predictions by moving beyond single-model approaches to ensemble and machine learning (ML) strategies.

Methodology
The research utilized a multi-stage approach combining simulation and clinical validation:

1. Data Generation: A virtual patient population was simulated based on cohorts from four published adult amoxicillin PopPK models. Steady-state trough concentrations were generated using cohort-specific dosing regimens to create model-development datasets.
2. Reference Methods: The study implemented established techniques as baselines, including Weighted Model Ensembling (WME) and Classification Tree (CT)-informed ensembling.
3. Novel Ensembling Strategies: Two new methods were developed:
   • Regression Tree (RT)-informed ensembling: Utilized RTs to predict the log individual prediction/observation ratio to assign model weights.
   • Factor Analysis of Mixed Data (FAMD): Assigned model weights based on the similarity of a patient's covariates to the original model cohorts.
4. Machine Learning Approaches: Four ML algorithms—Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), and XGBoost—were trained to predict the specific dose required to achieve target concentrations, utilizing patient covariates and dosing schemes as inputs.
5. Validation: All approaches were compared against single-model PopPK dosing, standard dosing, and a nomogram. Performance was evaluated on simulated data and externally validated using clinical data, including patients receiving continuous infusion.

Key Contributions

• Development of Novel Ensembling Techniques: The introduction of RT-informed ensembling and FAMD to dynamically weight PopPK models based on patient-specific data rather than static assumptions.
• Application of ML for Dose Prediction: The successful training of diverse ML algorithms to directly predict target-achieving doses, effectively bypassing the need for explicit model selection.
• Comparative Framework: A comprehensive evaluation demonstrating that ensemble and ML methods outperform traditional single-model and standard dosing strategies in the context of amoxicillin a priori dosing.

Results

• Performance on Simulated Data: Both ensembling methods (30–42% correct predictions) and ML approaches (36–39%) significantly exceeded single-model PopPK approaches (14–32%).
• Performance on Clinical Data: The novel RT-informed and FAMD ensembling strategies improved prediction accuracy by 6–10% compared to uninformed ensembling.
• Continuous Infusion Context: In clinical patients receiving continuous infusion, ensembling methods further enhanced performance, with FAMD achieving 49% correct predictions.
• General Efficacy: Most MIPD methods evaluated in the study outperformed standard dosing protocols.

Significance and Claims
The paper claims that ML-based ensembling and novel model weighting strategies eliminate the need for manual model selection while increasing the rate of target attainment for amoxicillin in ICU settings. The study posits that these approaches provide a more robust framework for a priori precision dosing, enabling faster therapeutic target achievement without the delay of concentration measurements. The authors maintain a modest stance, presenting these methods as improvements over existing single-model and standard practices rather than a definitive solution to all pharmacokinetic variability.