In therapeutic drug monitoring of vancomycin (VCM), the area under the concentration-time curve (AUC) is related to clinical efficacy and toxicity. Determining the maintenance for patient is necessary since VCM concentrations are affected by factors such as renal function. We constructed a machine learning-based model to estimate the maintenance dose to target an AUC of 400-600 mg⋅h/L in each combination of patient's factors. This retrospective observational study was conducted at two hospitals. Patients who received VCM intravenously with measured trough and another point (e.g., peak) concentrations within the November 2011 to March 2019 period were enrolled. We extracted the factors that affect VCM concentration and constructed a decision tree model using a classification and regression tree algorithm. Of the 1380 patients, 822 were included. Training data were split up to four times and included 24 subgroups. The average corrected VCM daily doses ranged 17.6-59.4 mg/kg. Estimated glomerular filtration rate, age, and body mass index were selected as predictive variables that affected the recommended daily dose. In the validation data, our model had slightly higher proportions of AUC of 400-600 mg⋅h/L than other nomograms. However, our model was based only on limited patients. Thus, further clinical studies are needed to develop a general-purpose model in the future. We successfully constructed a model that recommends VCM maintenance daily doses with AUC of 400-600 mg⋅h/L for each combination of independent variables. Our model has the potential for application as a simple decision-making tool for medical staff.
Keywords: area under the concentration–time curve; decision tree analysis; machine learning method; therapeutic drug monitoring; vancomycin.