A Machine Learning Approach to Predict Interdose Vancomycin Exposure.

Area of Science:

• Pharmacometrics and Machine Learning
• Drug Concentration Monitoring
• Computational Pharmacology

Background:

• Estimating vancomycin area under the curve (AUC) is challenging with discontinuous dosing.
• Machine learning (ML) offers a promising alternative to population pharmacokinetic (POPPK) models for AUC estimation.
• XGBoost algorithms can predict vancomycin AUC using early concentrations and patient data.

Purpose of the Study:

• To train XGBoost algorithms using simulated data to predict vancomycin AUC.
• To evaluate the performance of these ML algorithms against POPPK models in a real-world patient cohort.
• To assess the accuracy and precision of ML-based AUC estimation for vancomycin therapy.

Main Methods:

• Trained XGBoost algorithms on 6,000 simulations from six POPPK models.
• Evaluated algorithms using 2, 4, and 6 samples for AUC prediction, with resampling and external validation.
• Externally validated the 2-sample XGBoost model on 28 real patients, comparing it to POPPK model-based averaging.

Main Results:

• Trained algorithms demonstrated high accuracy in test sets: MPE/RMSE of 3.3/18.9% (2 samples), 2.8/17.4% (4 samples), and 1.3/13.7% (6 samples).
• External validation in real patients showed excellent performance for the 2-sample algorithm (MPE/RMSE < 1.5/12%).
• The 2-sample algorithm exhibited flexibility in sampling times and outperformed various POPPK approaches.

Conclusions:

• XGBoost algorithms trained on simulations accurately predict vancomycin AUC in real patients.
• This ML approach enhances confidence in AUC estimation when used alongside POPPK models.
• ML models offer a viable and precise alternative for vancomycin AUC monitoring.