Development and evaluation of uncertainty quantifying machine learning models to predict piperacillin plasma...

Area of Science:

• Pharmacology
• Machine Learning
• Critical Care Medicine

Background:

• Suboptimal beta-lactam antimicrobial concentrations are common in critically ill patients.
• Existing population pharmacokinetic (PopPK) models often lack predictive accuracy and uncertainty quantification for guiding dosing.
• There is a need for improved models to predict drug concentrations and adapt dosing regimens.

Purpose of the Study:

• To develop machine learning (ML) models for predicting piperacillin plasma concentrations.
• To incorporate uncertainty quantification into ML models for clinical applications.
• To evaluate the predictive accuracy and clinical utility of ML models compared to traditional PopPK models.

Main Methods:

• Prospective collection of piperacillin concentrations from critically ill patients receiving continuous infusion.
• Development of interpretable ML models using CatBoost and Gaussian processes.
• Implementation of a Quantile Ensemble method for uncertainty quantification with the CatBoost model.
• Evaluation using distribution coverage error on internal and external datasets.

Main Results:

• The CatBoost model achieved high predictive accuracy (RMSE 31.94-0.64 internally).
• The Quantile Ensemble method provided clinically useful individualized uncertainty predictions.
• Gaussian processes showed limitations in clinical applications due to their homoscedastic nature.

Conclusions:

• ML models can accurately estimate piperacillin concentrations with uncertainty quantification when using similar dosing regimens.
• Generalization to different dosing schemes remains a limitation.
• ML models show significant promise for integration into therapeutic drug monitoring programs.