A Clinical Data-Based Model for Drug Concentration Prediction of Vancomycin in Critical Patients

Background: Vancomycin is a glycopeptide antibiotic primarily used to treat infections caused by methicillin-resistant Staphylococcus aureus (MRSA). As a time-dependent antibiotic, the serum concentration of vancomycin is closely related to the clinical efficacy, toxicity and emergence of drug resistance. Therefore, therapeutic drug monitoring (TDM) is considered an important component of vancomycin treatment management. According to vancomycin surveillance guidelines, It is recommended to maintain a serum vancomycin concentration of 15-20 mg/L in patients with severe infections in order to improve clinical outcomes and prevent drug resistance. However, serum vancomycin concentration testing is not widely used in clinical practices, especially in resource-constrained areas and medical institutions, so individualized monitoring remains a challenge. Currently, studies on vancomycin concentration prediction generally use the population pharmacokinetic (PPK) model. However, this model is affected by many factors such as age, weight, and creatinine clearance rate. However, since critically ill patients have complex diseases accompanied by multiple organ dysfunction, vancomycin pharmacokinetics may be altered. In such patients, the evidence for concentration prediction using PPK models is insufficient.

Currently, the rapidly developing machine learning methods can help capture nonlinear variable relationships while making predictions through multiple variables to achieve a high degree of accuracy in prediction results. This study aims to use machine learning methods to establish an optimal model for predicting serum vancomycin trough concentrations in critically ill patients.

Objective: This study aims to extract the serum vancomycin concentration data from the Critical Care Database of Peking Union Medical College Hospital from January 2014 to December 2023 and use machine learning methods to establish the optimal model for predicting vancomycin concentrations in critically ill patients.

Methods: (1)This is a single-center, retrospective study. Data on serum vancomycin concentration in the Critical Care Database of Peking Union Medical College Hospital were screened. After meeting the eligibility criteria, the clinical data of included patients are collected through the inpatient medical record system, including demographic characteristics, severity scores, laboratory test information and treatment information. (2) After extracting the available data, five models of machine learning, including Linear Regression, Lasso Regression, Ridge Regression, Random Forest and LightGBM, are used to build prediction models. The model with the best prediction accuracy is selected based on the percent error (PE), the mean percentage error (MPE) and the mean absolute percentage error (MAPE). (3) The MIMIC-IV (Medical Information Mart for Intensive Care) database is used to conduct external validation of the model constructed by machine learning. Moreover, the investigators will compare the predictive performance of the PPK model with the constructed model.

Quality control: Patients who meet the inclusion criteria are included. Patients with missing information are not enrolled in order to reduce bias. The information of included patients is recorded and registered by a dedicated research person.