Pharmacometrics-based clustering for Pharmacoepidemiological risk stratification in real-world evidence for drugs with a narrow therapeutic index: a data science approach
A multidisciplinary project that aims at improving the safety/effectiveness of drugs with a narrow therapeutic index.
This project aims to develop an integrated approach for safety and effectiveness evaluation of drugs with a narrow therapeutic index (NTI) during their post-marketing phase that combines population pharmacokinetic-pharmacodynamic modelling and simulation from pharmacometrics, epidemiological reasoning from pharmacoepidemiology, and artificial intelligence techniques from data science. Pharmacometrics has mostly been used to optimize the trial design during clinical development by providing support for dose selection and the identification of intrinsic/extrinsic factors that warrant dose changes in patient subgroups. Irrespective of their complexity, modelling and simulation models used in pharmacometrics are intended to reduce the uncertainty about the benefits and risks of a drug, which overlap with the overall goals of pharmacoepidemiology that mostly operate during the post-marketing phase of a drug lifecycle. Despite the combination of these two disciplines may look intuitively reasonable, to date, evidence on the use of pharmacometrics modelling in combination with pharmacoepidemiology methodology for post-marketing drug safety and effectiveness surveillance is missing.
Pharmacoepidemiology may greatly benefit from pharmacometrics. Currently, the pharmacoepidemiological approach highlights clinically significant associations between certain drugs and scenarios of adverse events (including reduced efficacy), altered drug utilization patterns, such as discontinuation and treatment switching as well as increased clinical event rates associated with both increased and decreased pharmacologic actions. Although methods to establish causal inferences can be used in this discipline, unmeasured confounding factors may produce bias, thus requiring a theoretical basis (biological plausibility) that can be provided by pharmacometrics. The hypothesis generated by the pharmacoepidemiologic approach can then be evaluated by integrating population pharmacokinetic-pharmacodynamic models to evaluate the biological, physiological, and drug-related causes of the observed reduced efficacy or increased risk of adverse event.
In this scenario, an additional improvement to such an integrative approach can be provided by data science and, in particular, by artificial intelligence. Artificial intelligence may provide support for pharmacometricians to identify intrinsic and extrinsic factors that warrant dose adjustment in patient subgroups from real-world data. This integration is highly needed considering that, to date, treatment effectiveness and safety estimates are generally estimated as the average of the overall study populations and little is known on the subgroups of patients for whom the treatment may be more beneficial or harmful, especially among those patients requiring dose adjustment. A variety of methods has been developed in pharmacoepidemiology and pharmacometrics for such analysis. The conventional generalized linear model can include prognostic variables as the main effect and predictive variables in interaction with the treatment variable. A statistically significant and large interaction effect usually indicates potential subgroups that may have different responses to the treatment. However, the conventional regression method requires specification of the interaction term (e.g. combination of diseases and treatments), which requires knowledge of predictive variables. This conventional approach also becomes infeasible when there is a large number of comorbidities and co-treatments. A variety of artificial intelligence techniques have been developed for this purpose in a hypothesis-free setting. However, their usefulness has primarily been investigated in clinical trials and evidence is sparse for real-world cohorts.
Our integrated approach will consider the measurement of relevant population-level clinical outcomes rather than more discrete pharmacological changes that are not captured in secondary data sources (e.g. Danish administrative registers), e.g. increased rate of heart attacks vs. increased heart rate variability. Additionally, it will focus on drugs for which variation in plasma concentration may have clinically relevant consequences in terms of safety and efficacy. For these reasons, as stated above, the project will focus on NTI drugs that are defined according to the U.S. Food and Drug Administration (FDA) as “those drugs where small differences in dose or blood concentration may lead to dose and blood concentration-dependent, serious therapeutic failures or adverse drug reactions. Serious events are those which are persistent, irreversible, slowly reversible, or life-threatening, possibly resulting in hospitalization, disability, or even death”. Examples of NTI drugs include warfarin, levothyroxine, carbamazepine, digoxin, lithium carbonate, phenytoin, and theophylline.
The project spans three specific aims:
Based on these considerations, the objectives of this project are
1) to use population pharmacometric modelling and simulation techniques to cluster patients at high risk of dose modulation among those exposed to NTI in Danish administrative registers,
2) to compare the risk of clinically relevant outcomes (e.g. hospitalization and survival) among clusters and
3) to use artificial intelligence techniques to optimize clustering models developed by pharmacometricians.Our linking of missense variants to patient disease cohorts will be very valuable for: (i) more accurate patient sub-diagnoses for genotype-based personalised treatment, (ii) patient stratification upon entering clinical trials, and (iii) increased precision in delineation of changes in clinical therapy, (iv) personalise medicine prescriptions based on GPCR genotypes, (v) prioritise drugs for pharmacovigilance investigations, and (vi) design post-market follow-up studies e.g. drug repurposing.
Principal investigator (PI): Assist. Prof. Maurizio Sessa1
Co-principal investigator (Co-PI): Assist. Prof. Eva Sverrisdóttir2
Research team members: Prof. Morten Andersen1 and Assoc. Prof. Trine Meldgaard Lund2
1Pharmacovigilance Group, Department of Drug Design and Pharmacology, University of Copenhagen, DK-2100 Copenhagen, DK
2Pharmacometrics Group, Department of Drug Design and Pharmacology, University of Copenhagen, DK-2100 Copenhagen, DK