A multi-model averaging approach improves the performance of model-guided infliximab de-escalation in patients with inflammatory bowel diseases
Wannee Kantasiripitak (1), An Outtier (2,3), Debby Thomas (1), João Sabino (2,3), Sebastian G. Wicha (4), Séverine Vermeire (2,3), Marc Ferrante (2,3), Erwin Dreesen (1)
(1) Department of Pharmaceutical and Pharmacological Sciences, University of Leuven, Belgium (2) Department of Gastroenterology and Hepatology, University Hospitals Leuven, Belgium (3) Department of Chronic Diseases and Metabolism, University of Leuven, Belgium (4) Department of Clinical Pharmacy, Institute of Pharmacy, University of Hamburg, Germany
Objectives: Underexposure to infliximab is a common cause of loss of response in patients with inflammatory bowel diseases. Dosage regimen escalations (higher dose and/or frequency) are widely practised to boost infliximab trough concentrations (TCs) and restore the response. However, long-term maintenance of the escalated dosage regimen has financial, practical, and potentially safety implications, and is therefore not warranted. Since infliximab exhibits substantial interindividual pharmacokinetic variability, dosage regimen de-escalation could put patients at risk for underexposure and trigger again the loss of response. To ensure adequate – but not unnecessarily high – exposure, we aimed to identify the best population pharmacokinetic model or a combination of models for guiding infliximab de-escalation.
Methods: Data of 55 patients who underwent infliximab dosage regimen de-escalation after an earlier successful escalation were obtained from a retrospective, single-centre cohort study. One to three infliximab TCs (TC-2, TC-1, TC0) before initiation of the dosage regimen de-escalation – but not necessarily consecutive – were used in addition to covariate data to predict the next infliximab TC (TC+1). The predictive performance of a single-model approach using seven previously published infliximab population pharmacokinetic models (Aubourg 2015, Brandse 2017, Dotan 2014, Dreesen 2020, Fasanmade 2009, Fasanmade 2011, and Grišic 2020) was compared with multi-model approaches (a model selection algorithm [MSA] and a model averaging algorithm [MAA] [1]) using the models jointly. A weighting scheme based on the squared prediction errors was implemented in the multi-model approach. In addition to visual predictive checks, relative bias (rBias) and relative root mean square error (rRMSE) were used to determine accuracy and precision, respectively. An rBias between ±20% with a 95% confidence interval including zero was considered clinically acceptable. The rRMSE was to be as low as possible, but no threshold was pre-specified.
Results: Covariate-based (a priori) predictions with any model was clinically unacceptable (rBias -45% to +91%, rRMSE 57% to 160%). The predictive performances of all models greatly improved by considering at least one infliximab TC (Bayesian forecasting; TC0: rBias -17% to +20%, rRMSE 30% to 59%). Using additional previous TCs improved the predictions only marginally (TC-1 and TC0: rBias -19% to +13%, rRMSE 30% to 78%; TC-2, TC-1, and TC0: rBias -14% to +16%, rRMSE 27% to 49%). Three out of seven models (Aubourg, Dotan, and Dreesen) displayed clinically acceptable bias when using one to three TCs (rBias -4% to +7%).
Both MAA and MSA resulted in clinically acceptable predictions, with rBias +1% and +9%, respectively, when considering TC0 and rBias -2% and -3% when considering all three TCs. MAA performed systematically better than MSA, not only in terms of accuracy but also in terms of precision. Performance of the MAA was less sensitive to the number of TCs considered in Bayesian forecasting, while the predictive performance of the MSA and single-model approaches improved with the number of samples considered.
The Dreesen model had the highest weight in the majority of patients, irrespective of the number of TCs considered (41/55 [75%] of patients when using TC0 only; 30/55 [55%] of the patients when using all three TCs). Although the Dreesen model was built based on data from patients with Crohn’s disease, the model performed equally well in patients with ulcerative colitis (highest weight in 12/17 [71%] of patients when using TC0 only). Having the highest weight in most patients (which indicates the best fit to the available TCs) did not make the Dreesen model the least biased in the single-model prediction approach.
Conclusions: The multi-model approaches, especially MAA, provided a more reliable Bayesian forecast compared to the single-model approach for guiding infliximab de-escalation in patients with inflammatory bowel diseases. The Dreesen model displayed the highest weight in the multi-model approaches.
References:
[1] D. W. Uster et al. CPT (2021) 109, 175–18
