Evaluation of population pharmacokinetic model predictive performance in adults on hemodialysis treated with vancomycin
Dominic Tong1, Ron J Keizer1, Jasmine H Hughes1
1InsightRX
Introduction: The pharmacokinetics (PK) of vancomycin in patients undergoing hemodialysis is significantly influenced by both the hemodialysis process itself and the underlying physiological changes that necessitated it. As a front-line antibiotic, vancomycin exhibits considerable interindividual variability in PK. Rapid attainment of therapeutic exposure is associated with improved clinical outcomes, but overexposure is associated with nephrotoxicity [1-3]. Model-informed precision dosing (MIPD) tools can enhance patient care by optimizing vancomycin dosing and exposure at the bedside, but its effectiveness depends on a population PK (popPK) model that adequately describes patient PK for vancomycin under hemodialysis [4]. While many such popPK models have been published, no large-scale evaluation of their performance has yet been conducted. Objective: Evaluate predictive performance of popPK models for adult patients undergoing hemodialysis treated with vancomycin. Methods: We identified seven popPK models from literature that described vancomycin PK in adult patients undergoing hemodialysis. Deidentified retrospective data from 1246 patients treated across 1663 occasions totaling 5084 concentrations undergoing routine clinical care using InsightRX Nova were the basis of this study. At the point-of-care, clinicians chose an appropriate model for their patient and then either entered clearance representing HD or entered required covariates for a HD-specific model (Oda [5]). Data were available from two cohorts of patients: (1) patients undergoing intermittent hemodialysis (iHD); and (2) patients undergoing continuous renal replacement therapy (CRRT). All cohorts were tested on the Hui [6], Bae [7], Goti [8], modified Goti/Tong [9], Goti with hemodialysis (Goti HD), Ghouti-Terki [10], and several non-HD vancomycin models with a clinician-estimated HD clearance. In a subset of cohort 2, urine output and effluent flow rate measurements were available, so we also tested the Oda model (cohort 2A). Predictions were made iteratively using PKPDsim [11]; that is, data up to and including the first N samples were used to predict the (N+1)th sample. Each prediction was made using a MAP Bayesian estimation of pharmacokinetic parameters, which can be utilized to inform both initial dose selection and subsequent dose adjustments. Predictive performance was assessed on accuracy (within 2.5 mg/L or 20% of the observed value), normalized root mean square error (nRMSE), and mean percent error (MPE). Acceptable performance was set as 50%/70% (a priori / a posteriori) for accuracy, 50%/30% for nRMSE, and 50%/30% for MPE. Results: In cohort 1 (iHD), the Goti HD model performed best a priori while the Bae model performed best a posteriori. The Goti HD model achieved an accuracy of 63.2%/64.2% (a priori/a posteriori), an MPE of 2.8%/10.0%, and an nRMSE of 28.3%/35.3%, while for the Bae model, these metrics were 30.0%/70.2%, -26.7%/10.8%, and 47.0%/35.3%. In cohort 2 (CRRT), the Goti model, though developed on iHD patients, predicted best (a posteriori 54.3%/14.9%/38.6%), although performance in all models was generally lower than in the iHD cohort. However, the non-HD models with the clinician-estimated HD component (a posteriori 45.8%/-10.5%/34.7%) showed lower MPE and nRMSE but also lower accuracy. In cohort 2A, the Oda model was not the most accurate model, though it had the lowest MPE (-7.5%/-7.9%). The modified Goti model however had higher accuracy (37.9%/54.3% vs 30.2%/30.9%) and lower nRMSE (50.5%/38.6% vs 50.9%/47.2%). Conclusion: We found that performance in models used to predict vancomycin exposures for hemodialysis patients varies depending on whether patients were undergoing iHD or CRRT. No models met all performance thresholds for acceptable clinical performance, though models developed on larger, multi-center data sets generally outperformed more niche models. Use of detailed HD information, like with the Oda model, does not necessarily compensate for smaller, less varied data sets. Further work should focus on re-evaluating how HD information can be incorporated in popPK models, refitting existing models, and developing new models to incorporate clinician knowledge of hemodialysis in vancomycin PK.
[1] Cardile et al, Springerplus, 2015 [2] Casapao et al, AAC, 2015 [3] Rybak et al, J Pediatric Infec Dis Soc, 2020 [4] Keizer et al, CPT:PSP, 2018 [5] Oda K et al, Pharm Res, 2020 [6] Hui K et al, JAC, 2019 [7] Bae SH et al, Pharmaceutics, 2019 [8] Goti V et al, TDM, 2018 [9] Tong DMH et al, TDM, 2021 [10] Ghouti-Terki L et al, Nephron, 2017 [11] https://github.com/InsightRX/PKPDsim/
