Tingjie Guo (1, 2, 3), Reinier M. van Hest (2), Luca F. Roggeveen (1), Lucas M. Fleuren (1), Rob J. Bosman (4), Peter H.J. van der Voort (4), Armand R.J. Girbes (1), Ron A.A. Mathot (2), Paul W.G. Elbers (1), Johan.G. Coen van Hasselt (3)
(1) Department of Intensive Care Medicine, Amsterdam UMC, Location VUmc, The Netherlands; (2) Department of Pharmacy, Amsterdam UMC, Location AMC, The Netherlands; (3) Division of Systems Biomedicine and Pharmacology, Leiden Academic Centre for Drug Research (LACDR), Leiden University, The Netherlands; (4) Intensive Care Unit, OLVG Oost, Amsterdam, The Netherlands.
Introduction
Model-based dose adaptation (MBDA) based on empirical Bayes estimates (EBE) plays an important role in optimizing drug treatment in intensive care (ICU) patients. MBDA requires dosing histories and patient-specific covariates, which can be subject to human error. For instance, delays between electronic health records and actual dosing times are inevitable during routine patient care. A strength of MBDA is the possibility to include historical pharmacokinetic (PK) data, which may improve parameters estimation. This strength may turn into a weakness due to high PK variability in ICU patients. The aim of this study was to investigate the impact of: 1) delays in recorded versus actual dosing, and 2) inclusion of historical PK data from ICU patients, on the EBEs and associated dosing recommendations.
Methods
Vancomycin PK model
We used a published one-compartmental population PK model of vancomycin in ICU patients [1]. The model was validated in external datasets [2].
Dosing time error
Dosing time delay (DTD) was defined as the delay between recorded and actual time of dosing. We evaluated the impact of DTDs on EBEs and derived metrics to quantify its impact on treatment efficacy. A dataset was simulated (1000 patients) by sampling covariates (weight, CLcr) from a historical dataset [2]. A dose regimen of 1000 mg B.I.D was assumed. The simulated PK data was used to calculate EBEs for different DTDs (0, 0.25, 0.5, 1, 2 hours or sampled from a truncated normal distribution). We considered two scenarios: S1: sampling after the 1st dose; S2: sampling before the 4th dose. Sampling times for these scenarios were obtained according to the following designs: D1: trough; D2: mid-interval; D3: peak & trough; D4: peak & mid-interval; D5: peak, mid-interval & trough. For these simulation scenarios we generated recommendations of both loading dose (LD) and maintenance dose (MD) required to reach the efficacy target of AUC24h,ss≥400 mg*h/L. For these recommendations we computed the probability of target attainment (PTA), and a probability of contradicted dose decision (PCD), reflecting a contradicted direction of the dose recommendation (e.g. predicted dose reduction instead of dose increase).
Inclusion of historical data
To assess the value of historical data in predicting vancomycin concentrations in ICU patients, we calculated EBEs in a real dataset of 490 patients [2], which included PK data for multiple occasions. We generated splits in the dataset for each patient, predicting “future” left-out concentrations based on previously observed data. In a simulation study, we investigated the impact of including historical PK data assuming random inter-occasion variability or time-varying changes in PK parameters.
Results
Dosing time error
When including more than one sample, the EBEs for volume of distribution (V) showed clear bias (mean error (ME) <0.417) but for clearance (CL) were subtle (ME>-0.096) when DTD increased. When no DTD occurred, EBEs for CL and V were unbiased but imprecise (RMSE of both >0.2). When drawing samples before the 4th dose instead of after the 1st dose, the PCD of MD decreased while the PCD of LD increased. The average PCD was >10% for the MD in all scenarios. An increase in DTD of >0.5h can lead to the decrease of PTA (<15%) compared to when no DTD occurred. Nonetheless, the PTA was <60% in all evaluated scenarios suggesting the undertreatment for ICU patients despite adapted dose.
Inclusion of historical data
When including historical PK data for >3 days (and < 14 days) in the past, vancomycin concentrations were underpredicted (<20%). When only historical samples from the previous day were included, no bias was observed. A time-varying trend for CL was observed indicating deterioration in vancomycin CL not captured by the included covariate for renal function. Our simulation study confirmed that a time-varying decreases in CL can explain the observed behavior and quantified the impact in dose adjustment errors. In contrast, introducing random inter-occasion variability (20%) into the simulated data did not reproduce the bias of predicted concentration when including historical data. This finding underlines the importance of considering time-varying trends in PK parameters in ICU patients.
Conclusions
Errors in recorded versus actual dosing times lead to bias in EBEs and may impact treatment efficacy. Inclusion of historical PK data to support MBDA may negatively impact model-based dose recommendations.
References:
[1] Roberts JA et al. 2011. Vancomycin dosing in critically ill patients: Robust methods for improved continuous-infusion regimens. Antimicrobial Agents and Chemotherapy 55:2704–2709.
[2] Guo T et al. 2019. External evaluation of population pharmacokinetic models of vancomycin in large cohorts of intensive care patients. Antimicrobial Agents and Chemotherapy. (Accepted)
Reference: PAGE 28 (2019) Abstr 9179 [www.page-meeting.org/?abstract=9179]
Poster: Clinical Applications