Marina Dimitriou 1, Siv Jönsson 2, Joakim Nyberg 2, Niclas Jonsson 2
1 Department of Pharmacy, Uppsala University (Uppsala, Sweden), 2 Pharmetheus AB (Uppsala, Sweden)
Introduction: Model-Informed Precision Dosing (MIPD) uses mathematical models to interpret therapeutic drug monitoring samples and predict personalized dosing strategies [1]. MIPD can also be used before any observation data is available, for a priori predictions, by utilizing the population model and patient-specific covariates to determine the initial dose required to achieve therapeutic/safety targets [1].
Full Random Effect Models (FREM) is an innovative covariate modelling method, robust to missing covariate data, as it treats covariates as observed data rather than error free predictors [2]. Missing covariate observations are treated as missing values of dependent variables, so they are implicitly handled by the model [2]. Additionally, FREM, compared to full fixed effects models with mean imputation, have been shown to predict typical individual values closer to the non-missing situation [3]. In an MIPD setting, FREM can be used to derive fixed effect models (FEM) for any subset of available covariates without parameter re-estimation and therefore enable individual dose predictions regardless of missing patient covariate data.
Objectives: To benchmark FREM performance against a reference FEM under ideal conditions (no missing covariates), as well as under different missing covariate scenarios for a priori dose predictions.
Methods: A one-compartment population pharmacokinetic model with linear elimination for linezolid [4], incorporating age, BSA, and eGFR as covariates on clearance and BSA on volume of distribution was used to simulate 24-hour pharmacokinetic profiles (24 observations) for 1000 individuals following a single 600 mg oral dose. Individual covariate data were derived from a cohort of the NHANES [5] with complete records for age, sex, weight, height, serum creatinine, eGFR, and BSA. A log-distributed dummy variable was also included as internal reference. Two models were fit to the simulated data: The reference FEM (REF) with exponential parameter-covariate relationships for age, BSA, and eGFR and FREM including all covariates. Individual covariate data for a new set of 1000 individuals were derived from the NHANES cohort. From this dataset the individual typical clearance values were derived and used for the calculation of a priori dose predictions. An Area Under the Curve to Minimum Inhibitory Concentration ratio (AUC0-24h/MIC) of approximately 100 is a minimum target commonly used for linezolid efficacy [6]. Assuming a bacterial infection with MIC of 1 mg/L, an AUC of 100 mg*h/L was set as a target and a priori dose predictions were calculated as Dose=Clearance*AUC. Predictions were generated using the REF with all covariate information available, as well as across various missing covariate scenarios using mean imputation. Similarly, from FREM, corresponding FEM were derived which included the subset of covariates used in each REF, as well as univariable FEM for the covariates that were not included in the REF and were all used for predictions. Root Mean Squared Relative Error (RMSE) was calculated for the dose predictions across all models, using predictions from REF with all covariate information available, as the “true” dose. Model simulation and estimation was performed using NONMEM (version 7.5.1) and Pearl-Speaks-NONMEM (version 5.6.1). Model predictions, plotting and post-processing of outputs were performed using R (version 4.2.2).
Results: FEM derived from FREM were superior to the corresponding REF across all missing covariate scenarios. The best performing model was FEM including BSA and eGFR, with an RMSE of 5%. Amongst the univariable models for the covariates included in REF, FEM eGFR was the best performing with an RMSE of 13% compared to 17.4% for REF. For the univariable FEM of covariates that were not included in REF, FEM including weight performed best with RMSE of 17%, which can be compared to RMSE of 28% for the REF with no covariate information.
Conclusion: The FREM derived models outperform REF in missing covariate data scenarios, while they can also utilize available covariates beyond those included in the reference model to generate more accurate a priori dose predictions. Therefore, these findings suggest that FREM provides a more robust and flexible method for a priori dose predictions in MIPD settings, where covariate availability can vary.
References:
[1] Minichmayr IK, Dreesen E, Centanni M, Wang Z, Hoffert Y, Friberg LE, Wicha SG. Model-informed precision dosing: State of the art and future perspectives. Advanced drug delivery reviews. 2024 Dec 1;215:115421.
[2] Yngman G, Bjugård Nyberg H, Nyberg J, Jonsson EN, Karlsson MO. An introduction to the full random effects model. CPT: Pharmacometrics & Systems Pharmacology.2022;11(2):149–60.
[3] Nyberg J, Jonsson EN, Karlsson MO, Häggström J. Properties of the full random-effect modeling approach with missing covariate data. Statistics in Medicine.2023 Dec 21;43(5):935–52.
[4] Crass RL, Cojutti PG, Pai MP, Pea F. Reappraisal of linezolid dosing in renal impairment to improve safety. Antimicrobial agents and chemotherapy. 2019 Aug;63(8):10-128.
[5] Centers for Disease Control and Prevention (CDC). National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey August 2021-August 2023 (2024), https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?Cycle=2021-2023
[6] Rao GG, Konicki R, Cattaneo D, Alffenaar JW, Marriott DJ, Neely M, IATDMCT Antimicrobial Scientific Committee. Therapeutic drug monitoring can improve linezolid dosing regimens in current clinical practice: a review of linezolid pharmacokinetics and pharmacodynamics. Therapeutic drug monitoring. 2020 Feb 1;42(1):83-92.
Reference: PAGE 34 (2026) Abstr 12244 [www.page-meeting.org/?abstract=12244]
Poster: Methodology - Covariate/Variability Models