This document summarizes a presentation on parametric modelling of time to onset of adverse drug reactions using parametric survival distributions. It discusses using hazard functions from parametric distributions to analyze reported time to onset data from EudraVigilance to detect safety signals. Two examples are analyzed: liver injuries with bosentan and infections with TNF-alpha inhibitors. The results show patterns consistent with the known safety profiles and mechanisms of the drugs. Future work could improve understanding of reporting mechanisms and test the approach on more data.
The masking effect of measures of Disproportionality Analysis
Parametric Modelling Time To Onset
1. Parametric modelling of time to onset
of adverse drug reactions using
parametric survival distributions
F. Maignen
Principal scientific administrator
European Medicines Agency
Plan of the presentation
1. Conflicts of interests and disclaimers
2. Background and rationale of the project
3. Materials and methods
4. Results
5. Interpretation and
discussion
6. Conclusions and
future directions
2. Conflicts of interest and more disclaimers
• P. Tsintis and M. Hauben have contributed to this study
• Other external contributions received from other Experts
• No external funding was received for this study
• I do not have any financial interests with the Pharmaceutical
industry or any IT software provider (declaration available from
the Agency)
• I thank the two Companies which have given their approval to
publish these results
• Disclaimer on the views expressed in this presentation wrt
European Medicines Agency
• No claim on a “better” safety profile on any medicinal product
mentioned in this work should be made.
Disclaimers (cont.)
ACKNOWLEDGEMENTS
• No external source of funding was used to perform this study. The
implementation of EudraVigilance was undertaken by the EudraVigilance team
at the EMEA lead by Dr Sabine Brosch. The following authors: FM has no
conflicts of interest with the pharmaceutical industry (declaration of interest
available from EMEA). P. Tsintis contributed to the study when he was working
for the EMEA. M. Hauben is also working in Department of Medicine, Risk
Management Strategy, Pfizer Inc., New York, New York University School of
Medicine, Departments of Community and Preventive Medicine and
Pharmacology, New York Medical College, Valhalla, New York, USA and for the
School of Information Systems, Computing and Mathematics, Brunel University,
London, England . None of the authors have any conflict of interests with any
statistical software provider. Valuable comments on this work were received
from Nils Feltelius, Hans-Georg Eichler, Francesco Pignatti, Xavier Kurz, Jim
Slattery and Anders Sundström.
DISCLAIMER
• The views expressed in this presentation are the personal views of the author(s)
and may not be understood or quoted as being made on behalf of or reflecting
the position of the European Medicines Agency or one of its committees or
working parties.
3. Background
Background
• The time to onset of adverse drug reactions is directly connected to the
underlying mechanism of the toxicity associated with a medicine (DoTS
classification)
• The current quantitative methods do not integrate any information concerning
the underlying toxic mechanism of the suspected medicinal product (some rare
studies conducted by A. Bate and E. Van Puijenbroek).
• Current methods used to analyse the reported time to onset of adverse drug
reactions in Pharmacovigilance
• Simple histograms (or LOESS)
– Only provide a partial view of the evolution of the risk
– The visualisation of the risk highly depends on the number of bins and
bandwidth
– Difficult to find a “risk window”
– Output can be awful (LOESS).
• Other non-parametric methods
– Kaplan-Meier estimate of the survivor function: can be difficult
to interpret and difficult to actually visualise the exact
evolution of the risk.
• Find patterns of toxicity (true signals) via the hazards
4. Rationale for the study: hazard and hazard
functions
• The hazard expresses the risk that something happens at a
certain time t (does not help a lot).
• The hazard function specifies the instantaneous rate at which
events / failures occur for items which survived until time t.
• Some recent classifications of adverse drug reactions (DoTS)
includes the time relatedness as one key elements the
classification.
• Therefore (in theory) the hazard should be directly connected to
the underlying mechanism of the toxic effect resulting in an
adverse drug reaction.
• Parametric survival distributions have a hazard function which is
specified by a function (in opposition to non-parametric methods
such as CPH).
Hazard fcts of parametric survival dist.
Kalbfleisch and Prentice. The statistical analysis
of failure time data. Second ed. Wiley and
sons.
5. Reported hazard of occurrence: a phenomenon
involving several mechanisms
• P(occur.)*P(diag./occur.)*P(rep./diag.)(1)
• P = prob. failure conditional on survival
until time t.
• Lim f(x)*g(x) = Lim f(x)*Lim g(x)
• Then when we take Lim t -> 0 (1)
becomes.
h(occur.)*h(diag./occur.)*h(rep./diag.)
PD Monitoring and Awareness
Toxicology profile “RM” activities Reporting mechanisms
Efficacy / duration tt Awareness
Materials and methods
6. Materials and methods
• We have used parametric survival distributions to perform a
modelling of the reported time to onset to compute and plot the
corresponding hazard functions for signal detection purposes (in
a broad sense).
• The objective is to illustrate (and better understand the elements
of interpretation of) the use of hazard functions for signal
detection purposes using two real examples.
• Study conducted on a spontaneous reported database
(EudraVigilance).
• Two examples have been used
in the study:
Liver injuries associated with Bosentan
Infections associated with the use
of TNF alpha inhibitors
Materials and methods
EudraVigilance reports
Computation of TtO
(> 5)
KM Fit parametric distribution
(Exp/Weibull/LogN/Normal)
Selection of best fit
Computation / plot haz fct
7. Criteria used to interpret the results
• The idea is to use a convergence of available
evidence together with the hazard functions
of the reported time to onset to assess
whether there is a signal:
– Existing signal
– Pharmacodynamic properties of the products
– Bradford-Hill criteria which have been used to
interpret the results of data mining algorithms.
Results
8. Liver injuries reported with bosentan
(descriptive stats)
Liver injuries reported with bosentan (KM)
9. Liver injuries reported with bosentan (result of the
fit of parametric distributions)
Liver injuries reported with bosentan
(hazard functions)
10. Bosentan – liver injuries
• Logical course of events some occurrences need
careful interpretation (blood bilirubin inc. and
[hyper]bilirubinemia)
• Pattern AST/ALT unusual for liver injuries (but not for
mitochondrial injuries from hepatocytes) but
consistent with clinical safety data
• Residual and constant risk of liver failure
• Consistent with the putative mechanism of toxicity
(dose-dpt)
• Consistent with the safety profile of bosentan (lack of
independence)
• Influence of the risk minimisation activities
Infections reported with the administration of TNF
alpha inhibitors (KM)
11. Infections reported with the administration of TNF
alpha inhibitors (hazard functions)
Reported risk of infection reported with the
administration of etanercept (hazard functions)
12. Infliximab and adalimumab (hazard functions)
TNF inhibitors - infections
• Striking similarities (early risk of UTI, sepsis,
pneumonia and herpes zoster) and differences
between products (TB and cellulitis)
• Consistent with the PD properties of the products and
results of clinical trials
• Differences could be explained by:
– PD/PK differences (half life of adalimumab significantly
longer than for the other two products, etanercept also binds
TNF beta, infliximab inhibits IFN gamma)
• Probable influence of RM activities / monitoring of the
patients (provided that the side-effect can be
detected / prevented - cf Bosentan).
13. Reported risk of tuberculosis reported with the
administration of TNF alphas
Risk of tuberculosis reported with the
administration of infliximab
14. TNF inhibitors - TB
• TNF alpha plays an important role in the control of
granulomatous infections
• Main difference is observed between infliximab / etanercept on
the one hand and adalimumab on the other
• Different PD properties between the products would implies
different profiles between infliximab and etanercept
• Different PK profile between adalimumab and the
other two products
• Awareness and risk minimisation: adalimumab
has been authorised after the first two products
when the risk of TB was established and
recommendations to monitor the patients had been
published (shift of risk of TB? Different reporting
pattern?).
Summary of the main results
• Pattern consistent with the logical course of action of the toxicity
(bosentan)
• Hazard consistent with the suspected mechanism of the toxicity
(bosentan – dose dependent)
• Consistency of the reported hazard of occurrence of the infections
across the 3 TNFs
• Consistency of the reported hazard of tuberculosis for infliximab
• Differences between the TNFs products (TB) could be explained by:
– Different pharmacological properties
– Different pharmacokinetic properties
– Different monitoring of the patients / reporting mechanisms
• Patterns consistent with the known safety profile of the product (two
analyses not completely independent)
• Hazards certainly influenced by awareness and risk management
activities
15. Discussion
Factors influencing the modelling
• Nplicates
– Method sensitive to duplication
like any other DMA
– Consider the cases of extreme
duplication
– Duplication vs clusters
• Data quality
– Accuracy of the dates
– Completeness and precision does not
mean accuracy
• Good documentation and FUp of the cases
16. Statistical issues and important limitations
• The work is still preliminary. Interpretation is
still based on explanations which involve
documented pharmacological or reporting
behaviours which can be subjective
• Issue with censoring and competing risks
• Absence of hypothesis testing +++
• Great difficulty to choose a suitable
comparator to build the hypothesis testing.
• Performances need to be tested
Statistical issues and important limitations
• Approach limited by the number and quality of the reports
• Influence of the reporting mechanisms +++
– since the modelling was performed on spontaneous
reporting data, the hazard does not have the usual
interpretation as an instantaneous probability of failure
conditional on survival to time t.
– As far as the spontaneous reports are concerned, the hazard
reflects a mixture of reporting behaviour and natural history
which cannot be disentangled.
• Multi-state model?
– A “complete model” would not be devoid of any limitations or would
rely on strong assumptions which may not be met.
– Spontaneous reporting does not collect all the information needed
to build such model).
– Situation dependent
17. CONCLUSIONS
Conclusions
• Encouraging work which illustrates the
potential use of hazard functions in signal
detection
• Inherent limitations of the spontaneous
reporting
• A lot of data manipulation
• Carefully consider the influence of reporting
mechanisms (biases) and data quality (cliché)
• Some statistical issues to be addressed
18. Future directions
• Better understand the reporting mechanisms
• Test the approach to discriminate true signals from
confounding (find negative examples)
• Build a test of hypothesis
• Use it in specific situations where a “shift” of the
hazard function could reflect an underlying /
intercurrent event
• Assess the performances on a larger scale of data
• Potentially able to disantangle the reporting
mechanisms by comparing functions from different
sources of collection of information
Acknowledgements
• Thank you to the persons who have
supported me in this work (list not limitative)
– Jim Slattery
– Xavier Kurz
– JM Dogne
– Anders Sundstrom
– Eugene Van Puijenbroek
– HG Eichler