Novel pharmacometric techniques to quantify and prevent iatrogenic withdrawal in children
Sebastiaan C. Goulooze (1), Erwin Ista (2), Monique van Dijk (2,3), Thomas Hankemeier (1), Dick Tibboel (2), Elke H.J. Krekels (1), Catherijne A.J. Knibbe (1,3)
(1) Department of Systems Biomedicine and Pharmacology, Leiden Academic Centre for Drug Research, Leiden University, Leiden, The Netherlands (2) Intensive Care and Pediatric Surgery, Erasmus Medical Center-Sophia Children’s Hospital, Rotterdam, The Netherlands (3) Division of Neonatology, Department of Pediatrics, Erasmus Medical Center-Sophia Children’s Hospital, Rotterdam, The Netherlands (4) Department of Clinical Pharmacy, St. Antonius Hospital, Nieuwegein, The Netherlands
Prolonged treatment with analgesics and sedatives can cause iatrogenic withdrawal syndrome (IWS) in children being weaned from these drugs (1). Reported incidences of IWS in the pediatric ICU are high and variable (5–87%), suggesting a need for both individualized weaning strategies and monitoring of IWS.
In the pediatric ICU of the Dutch Sophia Children’s Hospital, IWS monitoring relies on the SOSwithdrawal, a validated, objective scale that scores the presence of 15 withdrawal-associated symptoms (1,2). Some symptoms may however also be caused by pain, undersedation or delirium, complicating IWS monitoring. In addition to the SOSwithdrawal, the nurse also forms an expert opinion of withdrawal severity (2). In this study, the expert opinion was scored on a numeric rating scale (NRSwithdrawal) ranging from 0 (no withdrawal) to 10 (worst withdrawal), taking contextual factors such as the possibility of co-occurring pain, undersedation or delirium into account.
As a first objective, we aimed for a model-based quantification of the dynamics of drug dependence and withdrawal severity to ultimately individualize weaning strategies based on a child’s prior use of analgesics and sedatives. For this application, we developed a novel mechanism-based iatrogenic withdrawal model. This model was based on the NRSwithdrawal scores, as they provide more global albeit subjective information regarding withdrawal severity.
The second objective was to increase information obtained from the SOSwithdrawal scale by performing an item response theory (IRT) analysis on its item-level data. As the unidimensionality assumption is violated by the impact of pain, undersedation and delirium, regular IRT modelling was not be applicable (3) and we therefore developed two extensions of regular IRT: the supervised IRT (sIRT) (4), and supervised multi-dimensional IRT (smIRT). The rationale of these supervised IRT methods is to leverage the expert opinion contained in the NRSwithdrawal score to improve the quantification of withdrawal severity from the objective symptom data.
In 81 children (aged 1 month to 17 years), 1782 paired IWS assessments were performed with the SOSwithdrawal and NRSwithdrawal scales, during an observational clinical study (2). NRSwithdrawal scores range from 0 (no withdrawal) to 10 (worst withdrawal possible).
Characterizing the dynamics of NRSwithdrawal
A novel mechanism-based withdrawal model structure was developed to characterize the development and disappearance of drug dependence over time. The model contains hypothetical ‘dependence compartments’, which equilibrate with the central pharmacokinetic compartment at an estimated rate. Published population pharmacokinetic models were used in combination with individual dosing information to generate population predicted plasma concentration-time profiles in each patient of all key analgesics and sedatives (i.e. morphine, fentanyl, methadone, midazolam, lorazepam, propofol, esketamine and clonidine). Withdrawal severity was modelled using a linear relation with the drug deficiency, defined as the difference between the concentration in the ‘dependence compartment’ and the predicted concentration in the central compartment. A generalized truncated Poisson model with Markovian transition probability inflation was used to respect the bounded integer nature of the NRSwithdrawal (5). Using simulations, different weaning strategies were compared for different drugs.
Supervised IRT modelling of SOSwithdrawal items
Pharmacometric models based on item-level data of the SOSwithdrawal were developed using three IRT-based modelling techniques, i.e. regular IRT, sIRT and smIRT. For the sIRT and smIRT, the nurse’s NRSwithdrawal score was used as a ‘supervising variable’ to guide the latent variable of the model towards withdrawal (3). For the smIRT, one or two unsupervised latent variables were added to the sIRT model to limit violations of the local independence assumption, by accounting for the impact of conditions other than IWS that affect the SOSwithdrawal items.
To allow for a comparison of linear association between the NRSwithdrawal score and the latent variables of the regular IRT, sIRT and smIRT models, the parameters of the sIRT and smIRT models were fixed to their estimated values, and refitted to the data in the absence of the NRSwithdrawal scores, re-estimating only the distribution of a logit-normally distributed latent variable on the same 0–10 scale as the NRSwithdrawal scores. The AIC of linear models in which the total composite score of the SOSwithdrawal or the latent variable of a particular IRT model was the predictor, and the NRSwithdrawal score the dependent variable.
Using the mechanism-based withdrawal model, the dynamics of withdrawal and dependence could be established with statistical significance for fentanyl (p< 10-6), morphine (p=0.043) and esketamine (p=0.002). The estimated rate constant of the drug dependence compartment was higher for fentanyl (0.265 h-1) compared with esketamine (0.018 h-1) and morphine (0.008 h-1). As a result the dynamics of dependence for fentanyl are also affected by its clearance. For all drugs, the weaning period should be increased with increasing drug levels prior to weaning.
Compared with the total SOSwithdrawal score, the latent variable of the regular IRT model showed a weaker association with the NRSwithdrawal score (ΔAIC = +180.5). The re-estimated latent variables of the two supervised IRT models had a stronger association than the total SOSwithdrawal score, even when removing the NRSwithdrawal after estimation of the supervised IRT models, with the strongest association being observed with the smIRT with two latent variables (ΔAIC = -223.7). Interestingly, the residual item-pair correlations in the sIRT model corresponded with clinical knowledge regarding the SOSwithdrawal items that are associated with pain and undersedation, and these correlations were attenuated in the smIRT models.
The mechanism-based withdrawal model dynamically predicts IWS from fentanyl, morphine and esketamine and showed that the optimal strategy for weaning of drug-dependent children depends on both the type of drug and the drug levels prior to weaning.
For the SOSwithdrawal, where individual items are not only affected by withdrawal, regular IRT modelling was worse in terms of quantifying withdrawal, than analysis based on total SOSwithdrawal score. The quantification of withdrawal severity was improved when using sIRT and smIRT, in which the subjective NRSwithdrawal score was used to ‘guide’ the latent variable towards withdrawal. Using the supervised IRT models developed here to estimate the IWS severity from symptoms alone, can be useful when NRSwithdrawal scores are lacking, or as a supplement to the subjective NRSwithdrawal score.
 Ista et al. Intensive Care Med. 2009;35(6):1075-1081.
 Ista et al. Pediatr Crit Care Med. 2013;14(8):761-769.
 Ueckert. CPT PSP. 2018;7(4):205-218.
 Ide et al.Knowl Inf Syst. 2017;51:235-257.
 Plan et al. Clin Pharmacol Ther. 2012;91(5):820-828.