Development and external validation of a machine learning model to predict the initial dose of vancomycin for targeting an area under the concentration–time curve of 400–600 mg∙h/L

Abstract

Purpose

To develop and validate a novel artificial intelligence model for predicting the initial empiric dose of vancomycin, with the aim of achieving an area under the concentration–time curve (AUC) of 400–600 mg∙h/L, using individual clinical data.

Methods

Machine learning models were developed and validated internally and externally using data from adult patients who received intravenous vancomycin treatment between November 2020 and June 2023, using records from July to September 2023, sourced from two hospitals. This study included 205, 44, and 35 patients in the training, internal validation, and external validation sets, respectively. The Random Forest, XGBoost, and Ensemble models were evaluated based on the mean squared error, root mean squared error, R-squared value, and accuracy (±20 % of the actual dose).

Findings

In internal validation, the Random Forest model achieved the highest 20% Accuracy at 56.8%, while the XGBoost model achieved 56.8% and Voting Ensemble models attained 54.5% accuracy. In external validation, the XGBoost model achieved the highest 20% Accuracy at 74.3%, followed by Random Forest and Voting Ensemble, both also achieving 74.3% accuracy. The estimated glomerular filtration rate was the most significant predictor in all models, with body weight, serum creatinine, height, and age also prominently influencing the XGBoost model’s predictions.

Implications

The proposed model is capable of accurately predicting the optimal dose of vancomycin, which could aid physicians in making more informed treatment decisions regarding early therapy, particularly when AUC estimation systems are not accessible or readily available in routine clinical practice.