Example with IPD trial data

Valerie Aponte Ribero and Javier Sanchez Alvarez

July 26, 2022

Introduction
This document runs a discrete event simulation model using simulated individual patient data (IPD) to show how
the functions can be used to generate a model when IPD from a trial is available.

Packages and main options

library(descem)
library(dplyr)
library(tidyr)
library(survival)
library(survminer)
library(flexsurv)
library(flexsurvPlus)
library(purrr)
library(kableExtra)

options(scipen = 999)
options(tibble.print_max = 50)

Model concept
The model is represented below. All patients start in the progression-free state and may move to the progressed
state. At any point in time they can die, depending on the risk of each disease stage. Patients may also
experience a disease event which accelerates progression.
 Progression-accelerating event Progressed

Death

Progression-free

Generate dummy IPD trial data and fit survival

models
The dummy IPD trial data is generated below using the sim_adtte() function from the flexsurvPlus package.
Parametric survival models are fit to the dummy OS and TTP IPD. We are using the flexsurv package to fit the
parametric survival models.

#Generate dummy IPD

tte.df <- sim_adtte(seed = 821, rho = 0, beta_1a = log(0.6), beta_1b = log(0.6), beta_pd
= log(0.2))

#Change data frame to wide format

tte.df <- tte.df %>% select(-PARAM) %>% pivot_wider(names_from = PARAMCD, values_from =
c(AVAL,CNSR))

#Derive Time to Progression Variable from OS and PFS

tte.df <- tte.df %>% mutate(
AVAL_TTP = AVAL_PFS,
CNSR_TTP = ifelse(AVAL_PFS == AVAL_OS & CNSR_PFS==0 & CNSR_OS==0,1,CNSR_PFS),
Event_OS = 1-CNSR_OS,
Event_PFS = 1-CNSR_PFS,
Event_TTP = 1-CNSR_TTP
)

#Add baseline characteristics (sex and age) to time to event data

IPD <- tte.df %>% mutate(
SEX = rbinom(500,1,0.5),
AGE = rnorm(500,60,8)
)
#Plot simulated OS and TTP curves

#Overall survival
km.est.OS <- survfit(Surv(AVAL_OS/365.25, Event_OS) ~ ARMCD, data = IPD) #KM curve

OS.fit <- flexsurvreg(formula = Surv(AVAL_OS/365.25, Event_OS) ~ ARMCD, data = IPD, dist

= "Weibull") #Fit Weibull model to the OS data
OS.fit
#> Call:
#> flexsurvreg(formula = Surv(AVAL_OS/365.25, Event_OS) ~ ARMCD,
#> data = IPD, dist = "Weibull")
#>
#> Estimates:
#> data mean est L95% U95% se exp(est) L95% U95%
#> shape NA 1.1346 0.9764 1.3185 0.0870 NA NA NA
#> scale NA 3.1523 2.5170 3.9479 0.3620 NA NA NA
#> ARMCDB 0.5000 0.3066 0.0183 0.5949 0.1471 1.3588 1.0185 1.8129
#>
#> N = 500, Events: 150, Censored: 350
#> Total time at risk: 623.0253
#> Log-likelihood = -360.0964, df = 3
#> AIC = 726.1929

ggsurvplot(OS.fit, title="Overall survival",

legend.labs = c("Reference","Intervention"),
risk.table = TRUE)
#Time to progression
km.est.TTP <- survfit(Surv(AVAL_TTP/365.25, Event_TTP) ~ ARMCD, data = IPD) #KM curve

TTP.fit <- flexsurvreg(formula = Surv(AVAL_TTP/365.25, Event_TTP) ~ ARMCD, data = IPD, d

ist = "Weibull") #Fit Weibull model to the TTP data
TTP.fit
#> Call:
#> flexsurvreg(formula = Surv(AVAL_TTP/365.25, Event_TTP) ~ ARMCD,
#> data = IPD, dist = "Weibull")
#>
#> Estimates:
#> data mean est L95% U95% se exp(est) L95% U95%
#> shape NA 1.3757 1.2590 1.5033 0.0622 NA NA NA
#> scale NA 0.5865 0.5310 0.6478 0.0297 NA NA NA
#> ARMCDB 0.5000 1.0992 0.9277 1.2707 0.0875 3.0016 2.5286 3.5632
#>
#> N = 500, Events: 328, Censored: 172
#> Total time at risk: 358.9897
#> Log-likelihood = -269.0387, df = 3
#> AIC = 544.0774

ggsurvplot(TTP.fit, title="Time to progression",

legend.labs = c("Reference","Intervention"),
risk.table = TRUE)
Define DES model inputs
Inputs and variables that will be used in the model are defined below. We can define inputs that are common to all
patients ( common_all_inputs ) within a simulation, inputs that are unique to a patient independently of the
treatment (e.g. natural death, defined in common_pt_inputs ), and inputs that are unique to that patient and that
treatment ( unique_pt_inputs ). Items can be included through the add_item function, and can be used in
subsequent items. All these inputs are generated before the events and the reaction to events are executed.
Furthermore, the program first executes common_all_inputs , then common_pt_inputs and then
unique_pt_inputs . So one could use the items generated in common_all_inputs in unique_pt_inputs .
#Define variables that do not change on any patient or intervention loop
common_all_inputs <- add_item(
#Parameters from the survival models
OS.scale = as.numeric(OS.fit$coef[2]),
OS.shape = as.numeric(OS.fit$coef[1]),
OS.coef.int = as.numeric(OS.fit$coef[3]), #Intervention effect

TTP.scale = as.numeric(TTP.fit$coef[2]),
TTP.shape = as.numeric(TTP.fit$coef[1]),
TTP.coef.int = as.numeric(TTP.fit$coef[3]), #Intervention effect

#Utilities
util.PFS = 0.6, #Utility while in progression-free state
util.PPS = 0.4, #Utility while in progressed state
disutil.PAE = -0.02, #One-off disutility of progression-accelerating event

#Costs
cost.drug.int = 85000, #Annual intervention cost
cost.drug.ref = 29000, #Annual cost of reference treatment
cost.admin.SC = 150, #Unit cost for each SC administration
cost.admin.oral = 300, #One-off cost for oral administration
cost.dm.PFS = 3000, #Annual disease-management cost in progression-free state
cost.dm.PPS = 5000, #Annual disease-management cost in progressed state
cost.ae.int = 2200, #Annual adverse event costs for intervention
cost.ae.ref = 1400, #Annual adverse event costs for reference treatment
)

#Define variables that do not change as we loop through interventions for a patient
common_pt_inputs <- add_item(
#Patient baseline characteristics
Sex = as.numeric(IPD[i,"SEX"]), #Record sex of individual patient. 0 = Female; 1 =Male
BLAge = as.numeric(IPD[i,"AGE"]), #Record patient age at baseline

#Draw time to non-disease related death from a restricted Gompertz distribution

nat.death = draw_resgompertz(1,shape=if(Sex==1){0.102}else{0.115},
rate=if(Sex==1){0.000016}else{0.0000041},
lower_bound = BLAge) # Baseline Age in years
)

#Define variables that change as we loop through treatments for each patient.
unique_pt_inputs <- add_item(
fl.int = 0, #Flag to determine if patient is on intervention. Initialized as 0, but w
ill be changed to current arm in the Start event.
fl.prog = 0, #Flag to determine if patient has progressed. All patients start progress
ion-free
fl.ontx = 1, #Flag to determine if patient is on treatment. All patients start on trea
tment
fl.PAE = 0, #Flag to determine if progression-accelerating event occurred
 pfs.time = NA #Recording of time at progression
)

Events
Add Initial Events
We define now the possible events that can occur for the intervention and reference arm respectively using the
add_tte() function. Only patients in the intervention arm can have a treatment discontinuation, while patients in
both arms can have a progression, progression-accelerating and death event. The seed argument is being used in
the draw_tte() function which uses the i item to ensure that event times specific to each patient can be
replicated and updated at later time points.
 init_event_list <-
#Events applicable to intervention
add_tte(trt="int",
evts = c("Start",
"TxDisc",
"Progression",
"PAE",
"Death"),
input={
Start <- 0
Progression <- draw_tte(1,'weibull',coef1=TTP.shape, coef2= TTP.scale + TTP.coef.in
t,seed = as.numeric(paste0(1,i,simulation)))
TxDisc <- Inf #Treatment discontinuation will occur at progression
Death <- min(draw_tte(1,'weibull',coef1=OS.shape, coef2= OS.scale + OS.coef.int, see
d = as.numeric(paste0(42,i,simulation))), nat.death) #Death occurs at earliest of diseas
e-related death or non-disease-related death
PAE <- draw_tte(1,'exp',coef1=-log(1-0.05)) #Occurrence of the progression-accelerat
ing event has a 5% probability for the intervention arm
}) %>%

#Events applicable to reference treatment

add_tte(trt="ref",
evts = c("Start", #Events are the same except that there will be no treatment
discontinuation
"Progression",
"PAE",
"Death"),
input={
Start <- 0
Progression <- draw_tte(1,'weibull',coef1=TTP.shape, coef2= TTP.scale,seed = as.nume
ric(paste0(1,i,simulation)))
Death <- min(draw_tte(1,'weibull', coef1=OS.shape, coef2= OS.scale, seed = as.numeri
c(paste0(42,i,simulation))), nat.death) #Death occurs at earliest of disease-related dea
th or non-disease-related death
PAE <- draw_tte(1,'exp',coef1=-log(1-0.15)) #Occurrence of the progression-accelerat
ing event has a 15% probability for the reference arm
})

Add Event Reactions

Reactions for each individual event are defined in the following using the add_reactevt() function. Patients in
the intervention arm discontinue treatment at progression. Occurrence of the progression-accelerated event
results in an earlier progression (if it has not occurred yet). Note the use of the seed argument in the draw_tte()
function which ensures that the same seed is being used in the original and the updated draw of the time to
progression.
 evt_react_list <-
add_reactevt(name_evt = "Start",
input = {
modify_item(list(fl.int = ifelse(trt=="int",1,0)))
}) %>%
add_reactevt(name_evt = "TxDisc",
input = {
modify_item(list("fl.ontx"= 0))
}) %>%
add_reactevt(name_evt = "Progression",
input = {
modify_item(list("pfs.time"=curtime,"fl.prog"= 1))
if(trt=="int"){modify_event(list("TxDisc" = curtime))} #Trigger treatme
nt discontinuation at progression
}) %>%
add_reactevt(name_evt = "Death",
input = {
modify_item(list("curtime"=Inf))
}) %>%
add_reactevt(name_evt = "PAE",
input = {
modify_item(list("fl.PAE"= 1))

if(fl.prog == 0){ #Event only accelerates progression if progression ha

s not occurred yet
modify_event(list(
"Progression"=max(draw_tte(1,'weibull',coef1=TTP.shape, coef2= TTP.sc
ale + TTP.coef.int*fl.int, hr = 1.2, seed = as.numeric(paste0(1,i,simulation))),curtim
e))) #Occurrence of event accelerates progression by a factor of 1.2
}
})

Costs and Utilities

Costs and utilities are introduced below.

Utilities
Utilities relating to progression-free and progressed states are common across arms.
 util_ongoing <- add_util(evt = c("Start","TxDisc","Progression","Death","PAE"),
trt = c("int", "ref"), #common utility across arms
util = ifelse(fl.prog == 0, util.PFS, util.PPS)
)

util_instant <- add_util(evt = c("PAE"),

trt = c("int", "ref"), #common utility across arms
util = disutil.PAE
)

Costs
Costs are defined specific to each arm due to differences in drug, administration and AE costs. A one-off cost for
oral administration of the reference arm is also applied.

cost_ongoing <-
#Drug costs are specific to each arm
add_cost(
evt = c("Start","TxDisc","Progression","Death","PAE"),
trt = "int",
cost = (cost.drug.int +
cost.admin.SC * 12 + #Intervention is administered once a month
cost.ae.int) * fl.ontx +
ifelse(fl.prog == 0,cost.dm.PFS,cost.dm.PPS)
) %>%
add_cost(
evt = c("Start","TxDisc","Progression","Death","PAE"),
trt = "ref",
cost = cost.drug.ref + cost.ae.ref + #No ongoing administration cost as reference tr
eatment is oral
ifelse(fl.prog == 0,cost.dm.PFS,cost.dm.PPS)
)

#One-off cost only applies to oral administration of reference treatment, applied at sta
rt of treatment
cost_instant <- add_cost(
evt = "Start",
trt = "ref",
cost = cost.admin.oral
)

Model
Model Execution
The model is executed with the event reactions and inputs previously defined for each patient in the simulated
data set.
 #Logic is: per patient, per intervention, per event, react to that event.
results <- RunSim(
npats=as.numeric(nrow(IPD)), # Simulating the number of patients for whic
h we have IPD
n_sim=1, # We run all patients once (per treatment)
psa_bool = FALSE, # No PSA for this example
trt_list = c("int", "ref"),
common_all_inputs = common_all_inputs,
common_pt_inputs = common_pt_inputs,
unique_pt_inputs = unique_pt_inputs,
init_event_list = init_event_list,
evt_react_list = evt_react_list,
util_ongoing_list = util_ongoing,
cost_ongoing_list = cost_ongoing,
cost_instant_list = cost_instant,
ncores = 2,
drc = 0.035, # discount rate for costs
drq = 0.035, # discount rate for QALYs
input_out = c("BLAge","Sex","nat.death","pfs.time")
)
#> [1] "Simulation number: 1"
#> [1] "Time to run iteration 1: 2.36s"
#> [1] "Total time to run: 2.36s"

Post-processing of Model Outputs

Summary of Results
Once the model has been run, we can use the results and summarize them using the summary_results_det to
print the results of the deterministic case. The individual patient data generated by the simulation is recorded in
the psa_ipd object.

summary_results_det(results$final_output) #will print the last simulation!

#> int ref
#> costs 112575.87 101195.77
#> lys 3.72 2.87
#> qalys 1.71 1.23
#> ICER NA 13407.80
#> ICUR NA 24146.54

psa_ipd <- bind_rows(map(results$output_psa, "merged_df"))

psa_ipd[1:10,] %>%
kable() %>%
kable_styling(bootstrap_options = c("striped", "hover", "condensed", "responsive"))
evtname evttime cost qaly ly pat_id trt Sex BLAge nat.death pfs.tim

Start 0.0000000 0.000 0.0000000 0.0000000 1 int 0 64.43993 15.07995 NA

PAE 0.2684478 24583.505 0.1603272 0.2672120 1 int 0 64.43993 15.07995 NA

Progression 0.9153383 58315.562 0.3803189 0.6338648 1 int 0 64.43993 15.07995 0.915338

TxDisc 0.9153383 0.000 0.0000000 0.0000000 1 int 0 64.43993 15.07995 0.915338

Death 7.6975370 29307.951 2.3446361 5.8615902 1 int 0 64.43993 15.07995 0.915338

Start 0.0000000 0.000 0.0000000 0.0000000 2 int 0 65.07078 13.96292 NA

PAE 0.1375550 12625.167 0.0823380 0.1372301 2 int 0 65.07078 13.96292 NA

Progression 1.3788510 111267.504 0.7256576 1.2094294 2 int 0 65.07078 13.96292 1.378851

TxDisc 1.3788510 0.000 0.0000000 0.0000000 2 int 0 65.07078 13.96292 1.378851

Death 2.4502583 5015.856 0.4012685 1.0031713 2 int 0 65.07078 13.96292 1.378851

We can also check what has been the absolute number of events per strategy.

trt evtname n

int Death 500

int Start 500

int PAE 418

int Progression 398

int TxDisc 398

ref Death 500

ref Start 500

ref Progression 450

ref PAE 405

Plots
We now use the simulation output to plot the Kaplan-Meier curves of the simulated OS and PFS against the
observed Kaplan-Meier curves. The simulated progression-free survival curve is lower than the observed due to
the addition of the progression-accelerating event into the model.
#Overall survival
KM.death <- psa_ipd %>% filter(evtname=="Death") %>% mutate(Event = 1)

sim.km.OS <- survfit(Surv(evttime, Event) ~ trt, data = KM.death)

km.comb <- list(

Observed = km.est.OS,
Predicted = sim.km.OS
)

ggsurvplot(km.comb, combine = TRUE,

title="Overall Survival",
palette=c("coral","turquoise","turquoise3","coral3"),
legend.labs=c("Observed - Ref","Observed - Int","Predicted - Int","Predicted
- Ref"),
linetype = c(2,2,1,1),
xlim=c(0,10), break.time.by = 1, censor=FALSE)
 #Progression-free survival
km.est.PFS <- survfit(Surv(AVAL_PFS/365.25, Event_PFS) ~ ARMCD, data = IPD)
KM.PFS.DES <- psa_ipd %>% filter(evtname=="Death") %>% mutate(evttime = ifelse(is.na(pf
s.time),evttime,pfs.time),
Event = 1)

sim.km.PFS <- survfit(Surv(evttime, Event) ~ trt, data = KM.PFS.DES)

km.comb <- list(Observed = km.est.PFS,

Predicted = sim.km.PFS)

ggsurvplot(km.comb,combine = TRUE,
title="Progression-free Survival",
palette=c("coral","turquoise","turquoise3","coral3"),
legend.labs=c("Observed - Ref","Observed - Int","Predicted - Int","Predicted
- Ref"),
linetype = c(2,2,1,1),
xlim=c(0,5), break.time.by = 1, censor = FALSE)

Developed by Valerie Aponte Ribero, Javier Sanchez Site built with pkgdown (https://pkgdown.r-lib.org/)
Alvarez. 2.0.6.