Abstract
Background: Vancomycin population pharmacokinetic (popPK) models are essential for precision dosing but can perform poorly in underrepresented populations, including younger adults and patients with low serum creatinine (sCr). These 2 populations show altered pharmacokinetics (PK) compared with the general population; younger patients exhibit faster clearance (CL), whereas low sCr often reflects body composition rather than renal function, leading to implausible CL estimates. Objectives: This study aimed to develop 2 vancomycin popPK models on large, real-world data sets and compare model predictive performance in historically underrepresented populations of individuals: young adults and individuals with low sCr. Methods: Two new vancomycin popPK models were developed using retrospective InsightRX Nova platform data. The model for young adults was trained on patients aged 18-35 years (n = 6080), whereas the low sCr model included adults of all ages with sCr values of less than 0.8 mg/dL (n = 2354). Predictive performance was evaluated using mean percent error, normalized root mean square error (nRMSE), and accuracy (within 15% or 2.5 mg/dL of observed concentrations). Predictive performance was compared across 5 established models using both holdout internal validation data sets and a large, broad adult population external validation data set. Results: Both models used 2-compartment structures with combined proportional and additive error. The young adult model incorporated fat-free mass (FFM) and estimated glomerular filtration rate via Chronic Kidney Disease Epidemiology Collaboration 2021, whereas the low sCr model used FFM, raw sCr, age, and sex. The young adult model showed superior accuracy and lower nRMSE across prediction types compared with established models in the internal validation data set. For modeling low sCr, application of rounding sCr rules did not improve model performance. External validation on 384,876 patients confirmed both models generalized well beyond their development populations, with the young adult model performing across age groups and the low sCr model maintaining performance for sCr of less than 1.0 mg/dL. Conclusions: Developing population-specific vancomycin popPK models improves predictive performance for underrepresented groups. Using measured sCr without imposing clinically common sCr rounding rules optimizes performance. These findings support continued refinement of targeted PK models.