Experiences in applied clinical pharmacometrics: challenges, recommendations, and research opportunities
Ron Keizer, Ranvir Mangat, Srijib Goswami
InsightRX
Background: Pharmacometric research has produced a wealth of advanced modeling techniques and models that could be applied in clinical practice for model-informed precision dosing (MIPD) at the point-of-care. However, while pharmacometrics has gained a stronghold in drug development, the adoption and implementation of MIPD into clinical practice at the point of care is still remarkably low[1]. Arguably, the low uptake of MIPD has been hampered primarily by the lack of easy-to-use user interfaces and connections to electronic hospital records (EHR). Many scientific questions relevant for the effective translation of pharmacometric research into clinical decision support (CDS) tools do however still remain. An overview is presented here of practical challenges, recommendations, and opportunities for future research to optimizing the translation of pharmacometric knowledge into MIPD/CDS solutions.
Methods: Experiences from three years of developing and implementing MIPD/CDS tools into multiple hospitals and for various drug classes (antibiotics and chemotherapeutics) were collected, summarized, and used to identify knowledge gaps in pharmacometric research. Recommendations and potential solutions are presented for each gap.
Results: Challenges and gaps were identified in the following areas:
- Model selection: the identification of optimal models and clinical validation of models for a specific population is challenging and laborious but a necessary step for effective and safe translation. It is highly recommended that predictive power of the intended model(s) is evaluated before implementation in any new population or new clinical setting, and continuously monitored afterwards. Dedicated diagnostic tools should be developed to support this.
- Model updating: data collected by the MIPD tool provides insight into the predictive ability of the used models, but should also be used to update the model priors and/or model structure afterwards. (Semi-)automated updating of models and priors to better match a specific population is a largely unstudied area in pharmacometrics but could have large potential.
- Curation of EHR data: dosing and biomarker data imported from the EHR will inherently contain errors, missing data, and outliers. Algorithms can be developed and employed to filter and cleanse such data, although manual curation is still often necessary to extract optimal use.
- Handling of outlier patients: both parametric (e.g. flattening of priors) and non-parametric (e.g. extended grid) approaches allow outlier patients to be better captured by a model, although such strategies can be subjective and ad hoc. Retrospective and prospective evaluation of such strategies is highly recommended.
- Handling of interoccasion variability (IOV): while IOV is almost always relevant in the clinical setting and prone to induce bias if neglected [1], population PK models presented in literature do not always include IOV. Simulation studies indicate that use of individual estimates specific to a previous occasion lead to reduced predictive power in forecasting future exposure [2]. Selective weighting of more recent data points could reduce bias and improve predictive power. Strategies for handling IOV should be studied in analyses of retrospective datasets as simulation studies are necessarily limited by their assumptions.
- Use of historical data in individualized treatment: long term data may be available for a subject, possibly from multiple prior hospital visits. Incorporation of down-weighted historical data likely provides better predictive ability than neglecting or fully weighting them.
- Exposure-outcome relationships: while TDM software historically has focused on optimizing drug exposure (or pharmacodynamic measures), linking drug exposure to outcome should be the ultimate aim and will spur adoption of MIPD tools based on clinical and pharmacoeconomic considerations. Such links are currently sparse but can often be studied from routinely collected clinical data.
Conclusions: Challenges and knowledge gaps were identified regarding the optimal implementation population PK/PD models into MIPD/CDS solutions. Some gaps can be addressed by simulation studies or retrospective analysis of large retrospective datasets, while others will require dedicated prospective trials.
References:
[1] Darwich AS et al. Clin Pharmacol Ther. 2017 May;101(5):646-656
[2] Karlsson, MO & Sheiner, LB. J Pharmacokinet Biopharm. 1993 Dec;21(6):735-50.
[3] Abrantes JA et al. PAGE 26 (2017) Abstr 7290 [www.page-meeting.org/?abstract=7290]
