Sepsis is a life-threatening syndrome with highly heterogeneous and dynamic physiology that undermines the reliability of one-size-fits-all therapies. Besides source control, timely antimicrobial therapy remains the only causal treatment option for many patients. Vancomycin is commonly used to treat Gram-positive infections in critically ill patients, but dosing is challenging because of pronounced pharmacokinetic variability and interindividual heterogeneity in sepsis. Underdosing can promote antimicrobial resistance, while overdosing increases risk of toxicity. Predicting initial serum concentrations before therapeutic drug monitoring results are available could enable more precise, individualized dosing decisions.
This investigation was a single-center retrospective study of adult patients with sepsis admitted to an intensive care unit. Eligible patients received a vancomycin loading dose followed by continuous infusion and had at least one measured serum vancomycin concentration. The work was conducted using routinely collected data extracted from the hospital electronic medical record.
The study used routinely collected clinical health record data from the institution's electronic medical records. Inclusion criteria required adult sepsis patients who were administered a vancomycin loading dose and continued on a continuous infusion regimen, with documented serum concentrations available to serve as model targets. The abstract does not report the exact sample size, specific variable list, or timing of concentration measurements; those details were not reported in the abstract.
Three machine learning approaches were developed to predict the initial vancomycin serum concentration: elastic net regression, random forest, and XGBoost. To minimize bias and improve generalizability, model training, hyperparameter tuning, and performance evaluation used a stratified nested cross-validation scheme. This nested approach separates hyperparameter optimization from test-set evaluation to provide less biased estimates of predictive performance.
Model performance was compared against seven commonly used population pharmacokinetic (PK) models. The comparison sought to determine whether data-driven machine learning models built from routinely available clinical variables could outperform established PK approaches for predicting early vancomycin concentrations in septic ICU patients.
Among the tested models, the best-performing machine learning solution was the elastic net model. The elastic net model achieved an average root mean squared error (RMSE) of 6.19 for predicted initial vancomycin serum concentrations. By comparison, the best-performing traditional pharmacokinetic model produced an average RMSE of 7.83. These results indicate that the elastic net model reduced prediction error relative to conventional PK models in this dataset and setting.
Final analysis of the selected model identified several variables that substantially influenced predictive performance. Administration of noradrenaline emerged as an important clinical predictor, along with classical pharmacokinetic parameters such as body weight, serum creatinine, and the presence of chronic kidney disease. The influence of vasoactive therapy (noradrenaline) highlights the relevance of hemodynamic support and its potential effects on drug disposition in critically ill septic patients.
The machine learning approach described enables prediction of initial vancomycin serum concentrations prior to availability of therapeutic drug monitoring results. By integrating routinely recorded clinical variables, the model supports earlier, data-driven decision-making for individualized dosing in sepsis care. Improved early prediction accuracy may reduce risks associated with empiric under- or overdosing, potentially limiting resistance selection and toxicity.
However, the abstract does not report prospective clinical impact, specific dosing recommendations, or how model outputs would be integrated into clinical workflows. External validation, prospective evaluation, and assessment of clinical outcomes following model-guided dosing were not described in the abstract and therefore remain unresolved based on the provided text.
The SepsisDataNet.NRW study received ethics approval from the Ethics Committee of the Medical Faculty of Ruhr-University Bochum (registry No. 18-6606-BR). The work was supported by the Core Unit Bioinformatics at the Medical Faculty of Ruhr University Bochum (CUBiMed.RUB). The authors declared no competing interests. The study is single-center and retrospective; the abstract does not report sample size, data missingness handling, or external validation, which limits assessment of generalizability from the abstract alone.
Using routinely available electronic health record data from septic ICU patients, a machine learning-based elastic net model predicted initial vancomycin serum concentrations with lower RMSE than conventional population pharmacokinetic models (6.19 vs 7.83). Important predictors included noradrenaline administration, body weight, serum creatinine, and chronic kidney disease. The findings suggest that machine learning can support more precise, individualized vancomycin dosing prior to therapeutic drug monitoring results, but further external validation and prospective evaluation are needed to establish clinical benefit.