The goal of the jmpost
package is to fit joint models involving:
- a parametric time-to-event sub-model,
- a nonlinear (or linear) mixed-effect sub-model, describing individual time profiles (i.e. trajectories) for a continuous marker,
- a link function (a.k.a. association term).
More specifically, the model implemented in this package utilizes a modelling framework described previously [1-3] to link overall survival to tumour size data in oncology clinical trials.
[1] Tardivon et al. Association between tumour size kinetics
and survival in patients with urothelial carcinoma treated with
atezolizumab: Implications for patient follow-up. Clin Pharm Ther,
2019.
[2] Kerioui et al. Bayesian inference using Hamiltonian
Monte-Carlo algorithm for nonlinear joint modelling in the context of
cancer immunotherapy. Stat in Med,
2020.
[3] Kerioui et al. Modelling the association between biomarkers
and clinical outcome: An introduction to nonlinear joint models. Br J
Clin Pharm, 2022.
The models are implemented in STAN, and the
package provides a flexible user interface. Please reach out to us via
issues or email (see the DESCRIPTION
file) if you have comments or
questions or would like to get involved in the ongoing development,
thank you!
GitHub
You can install the current development version from GitHub with:
if (!require("remotes")) {
install.packages("remotes")
}
remotes::install_github("genentech/jmpost")
Please note that this package requires
cmdstanr
.
CRAN
This package has not been published to CRAN yet.
See also the model fitting vignette for more details. Here we present a very basic example here.
First we simulate a data set. In practice you want to follow a similar
structure of the input data and use DataJoint()
to bring it into the
right format.
library(jmpost)
#> Registered S3 methods overwritten by 'ggpp':
#> method from
#> heightDetails.titleGrob ggplot2
#> widthDetails.titleGrob ggplot2
set.seed(321)
sim_data <- SimJointData(
design = list(
SimGroup(50, "Arm-A", "Study-X"),
SimGroup(50, "Arm-B", "Study-X")
),
longitudinal = SimLongitudinalRandomSlope(
times = c(1, 50, 100, 150, 200, 250, 300),
),
survival = SimSurvivalWeibullPH(
lambda = 1 / 300,
gamma = 0.97
)
)
#> INFO: 1 subject did not die before max(times)
joint_data <- DataJoint(
subject = DataSubject(
data = sim_data@survival,
subject = "subject",
arm = "arm",
study = "study"
),
survival = DataSurvival(
data = sim_data@survival,
formula = Surv(time, event) ~ cov_cat cov_cont
),
longitudinal = DataLongitudinal(
data = sim_data@longitudinal,
formula = sld ~ time,
threshold = 5
)
)
Then we specify the joint model, here we use a Generalized Stein-Fojo model for the longitudinal part, and a Weibull proportional hazards model for the survival part. The longitudinal model impacts the hazard via a term for the derivative and another term for the time-to-growth.
joint_model <- JointModel(
longitudinal = LongitudinalGSF(),
survival = SurvivalWeibullPH(),
link = Link(
linkDSLD(),
linkTTG()
)
)
Finally we can sample the parameters via MCMC from the underlying Stan model. Note that in a real application you will choose more warm up and sampling iterations.
mcmc_results <- sampleStanModel(
joint_model,
data = joint_data,
iter_sampling = 100,
iter_warmup = 100,
chains = 1,
parallel_chains = 1
)
To cite jmpost
please see
here.