An ensemble model for prediction of vancomycin trough concentrations in pediatric patients

Abstract

Purpose: This study aimed to establish an optimal model to predict vancomycin trough concentrations by using machine learning. Patients and Methods: We enrolled 407 pediatric patients (age < 18 years) who received vancomycin intravenously and underwent therapeutic drug monitoring from June 2013 to April 2020 at Xinhua Hospital affiliated to Shanghai Jiaotong University School of Medicine. The median (interquartile range) age and weight of the patients were 2 (0.63–5) years and 12 (7.8–19) kg. Vancomycin trough concentrations were considered as the target variable, and eight different algorithms were used for predictive performance comparison. The whole dataset (407 cases) was divided into training group and testing group at the ratio of 80%: 20%, which were 325 and 82 cases, respectively. Results: Ultimately, five algorithms (XGBoost, GBRT, Bagging, ExtraTree and decision tree) with high R2 (0.657, 0.514, 0.468, 0.425 and 0.450, respectively) were selected and further ensembled to establish the final model and achieve an optimal result. For missing data, through filling the missing values and model ensemble, we obtained R2=0.614, MAE=3.32, MSE=24.39, RMSE=4.94 and a prediction accuracy of 51.22% (predicted trough concentration within ±30% of the actual trough concentration). In comparison with the pharmacokinetic models (R2=0.3), the machine learning model works better in model fitting and has better prediction accuracy. Conclusion: Therefore, the ensemble model is useful for the vancomycin concentration prediction, especially in the population of children with great individual variation. As machine learning methods evolve, the clinical value of the ensemble model will be demonstrated in the clinical practice.