Multimodal data-driven predictive model for vancomycin serum concentrations in postoperative cardiac surgery patients under cardiopulmonary bypass: a single-center retrospective study

TL;DR

This study developed a predictive model to estimate vancomycin levels in heart surgery patients, helping doctors avoid harmful or ineffective doses.

Contribution

A novel multi-modal data-driven model for predicting vancomycin serum concentrations in postoperative cardiac surgery patients.

Findings

Creatinine clearance rate and CPB duration are key predictors of nephrotoxicity risk.

Age and eGFR are significant predictors of subtherapeutic concentration risk.

The models showed strong predictive performance and stability in cross-validation.

Abstract

This study analyzed the pharmacokinetic data of vancomycin in patients after cardiac surgery under cardiopulmonary bypass (CPB). It aimed to identify factors affecting vancomycin plasma concentrations and develop a multi-modal data-driven prediction model for these concentrations.

A retrospective study included hospitalized patients who underwent cardiac surgery under cardiopulmonary bypass, received intravenous vancomycin for postoperative infection, and had plasma concentration monitoring. Patient demographic and other relevant information was collected. A dual-risk model based on the therapeutic window (10–20 mg/L) was established: ① nephrotoxicity risk group (≥20 mg/L vs. < 20 mg/L, n = 350); ② subtherapeutic concentration risk group (<10 mg/L vs. 10–20 mg/L, n = 211). Missing data were imputed using the K-nearest neighbor algorithm (KNN, K = 10) to ensure data integrity. Univariate logistic regression (α = 0.10) was used for initial variable selection. LASSO regression with 10-fold cross-validation selected optimal features, and a multivariate bidirectional stepwise regression (forward-backward method, AIC criterion) built the prediction model. Model validation used forest plots, nomograms, ROC curves, calibration curves, and decision curve analysis (DCA). Generalization performance was assessed via ten-fold cross-validation.

Our models demonstrated strong predictive performance for both nephrotoxicity and subtherapeutic concentration risks. In the nephrotoxicity model, creatinine clearance rate (CCr) (OR = 0.975, 95% CI: 0.965–0.984; P < 0.001) and CPB duration (OR = 1.200, 95% CI: 1.068–1.350; P = 0.002) were identified as independent predictors, achieving an AUC of 0.733 (95% CI: 0.680–0.787). For the subtherapeutic concentration risk model, age (OR = 1.035, 95% CI: 1.007–1.064; P = 0.016) and estimated glomerular filtration rate (eGFR) (OR = 0.981, 95% CI: 0.964–0.998; P = 0.032) were significant predictors, with an AUC of 0.758 (95% CI: 0.692–0.824). Both models showed good calibration (P > 0.05) on the Hosmer-Lemeshow test, with better discrimination in the low concentration risk group. DCA confirmed superior clinical net benefit for both models over full/intervention strategies. In ten-fold cross-validation, the AUC fluctuation was <5%, indicating good stability.

These models effectively predict steady-state vancomycin concentrations in post-CPB patients. Lower CCr and longer CPB duration increase nephrotoxicity risk, while younger age and higher eGFR were associated with increased subtherapeutic concentration risk. These findings facilitate early identification of high-risk patients by clinicians, enabling timely intervention for optimized dosing regimens.

Figures