PD Data Sciences, Methods, Collaboration, and Outreach Group (MCO), Roche Basel

03 Aug 2023

Acknowledgments

I thank Marcel Wolbers and Godwin Yung for a careful review an earlier version of this document and helpful comments.

Purpose

This tutorial outlines how one can predict an interim (IA) or final analysis (FA) clinical cutoff date (CCOD) in an ongoing trial, i.e. in a trial where

patients have already been recruited but have not yet experienced the event and/or
patients have already been recruited and have already experienced the event and/or
further patients will be recruited in the future.

Methodology has been implemented in the R package eventTrack (Rufibach (2021)). Please note that the R code included in this document has not been formally validated.

Resources and how to cite eventTrack

Citation of the R package:

citation("eventTrack")

## 
## Um Paket 'eventTrack' in Publikationen zu zitieren, nutzen Sie bitte:
## 
##   Rufibach K (2023). _eventTrack: Event Prediction for Time-to-Event
##   Endpoints_. R package version 1.0.3.
## 
## Ein BibTeX-Eintrag für LaTeX-Benutzer ist
## 
##   @Manual{,
##     title = {eventTrack: Event Prediction for Time-to-Event Endpoints},
##     author = {Kaspar Rufibach},
##     year = {2023},
##     note = {R package version 1.0.3},
##   }
## 
## ACHTUNG: Diese Zitationsinformation wurde aus der DESCRIPTION-Datei
## automatisch generiert. Evtl. ist manuelle Nachbearbeitung nötig, siehe
## 'help("citation")'.

Citation of the statistical methods used in eventTrack: the implemented methodology for the point prediction is taken from Fang and Su (2011).
The confidence intervals have been presented in a talk (slides) within a BBS seminar, Rufibach (2016).

Challenge

For a RCT with time-to-event endpoint we have that:

A part of patients is already accrued, another part is still to be accrued.
We plan an IA or FA after a pre-specified number of events.

The questions we want to answer are:

When does this number of events happen?
Can we assign a confidence interval around that timepoint?
On a more strategic level: how can we manage the potentially large uncertainty of our point estimates in clinical teams (and higher up)?

Methodology

For simplicity, assume for now only blinded data pooling the treatment arms is available. What we are interested in is the expected number of events, \(m(T, S)\), at a given timepoint \(T\). This quantity depends, of course, on the timepoint \(T\) and on the survival function at which events happen. If these two quantities are known, we can predict \(m\) using the following formula:

\[ m(t, S) = {\color{blue}\sum_{h \in {\cal{I}}_1} 1} \ + \ {\color{brown}\sum_{j \in {\cal{I}}_2}P\Big(\text{event in }(t_j, T] | \text{no event up to }t_j\Big)} \ + \ {\color{red}\sum_{j \in {\cal{I}}_3} P\Big(\text{event in }(a_i, T]\Big)} \\ = {\color{blue}n} \ + \ {\color{brown}\sum_{j \in {\cal{I}}_2} \frac{S(t_j) - S(T)}{S(t_j)} 1\{T > t_j\}} \ + \ {\color{red}\sum_{j \in {\cal{I}}_3} \Big(1 - S(T - a_i)\Big) 1\{T > a_i\}}. \] This formula defines the following sets of patients:

\({\color{blue}{\cal{I}}_1}\): patients who already had the event, \(n\).
\({\color{brown}{\cal{I}}_2}\): patients who are censored at \(t_j\).
\({\color{red}{\cal{I}}_3}\): patients to be recruited at \(a_i\).

So, in order to compute \(m(T, S)\) we need to specify two things:

a projection of when the remaining patients will be recruited, i.e. we need to specify the numbers \(a_i\),
what to plug in for \(S\).

Ideally, for \(S\) we would like to use a nonparametric estimate, such as e.g. Kaplan-Meier. However, the estimator of \(S\) needs to be defined at least up to the timepoint \(T\), which typically is (much) later than the maximally observed event or censoring time at the cleaning milestone. This means we need an estimate of \(S\) that extrapolates beyond where we have data available. Fang and Su (2011) propose to use a hybrid parametric / non-parametric estimate. In a first step, change points in the survival function (or alternatively formulated, breaks in the piecewise constant hazard) are detected. Secondly, the survival function before the last change point is estimated non-parametrically via Kaplan-Meier and the tail beyond the last change point is estimated parametrically, i.e. based on an exponential distribution.

How to infer the changepoint via sequential testing is described in Fang and Su (2011), based on the method initially described in Goodman, Li, and Tiwari (2011). It works as follows:

Determine the maximal number of changepoints, \(K\). \(K = 5\) generally works well.
Estimate a piecewise exponential hazard function with \(K\) changepoints and hazard values \(\alpha_1, \ldots, \alpha_{K + 1}\). Cutpoints \(\tau_1, \ldots, \tau_K\) are estimated as well.
Start with the model with no change points and perform a hypothesis test to compare it with the model with one change point. If we fail to reject the null hypothesis, then we stop and conclude that the final model has no change points.
If we reject the null hypothesis, then we will continue on to the next hypothesis test; compare the model with one change point, to the model with two change points. If the null hypothesis is rejected we test the subsequent hypothesis. That is, we will test for the existence of \(\tau_k\) only after verifying the existence of \(\tau_{k-1}\). The algorithm is continued until we fail to reject a hypothesis.
To control for overfitting the significance level is adjusted using an \(\alpha\)-spending function. For a familywise error rate of \(\alpha\), Goodman, Li, and Tiwari (2011) propose to use the local significance levels \(\alpha^*_k, k = 1, \ldots, K\) defined as \[ \alpha^*_k = \alpha / 2^{k-1}. \] In order to find a parsimonious model, this spending function requires strong evidence for choosing a more complicated model, a model with more change points, over a simpler one. Since the spending function is decreasing, the test for each additional change point will be conducted at a more conservative local significance level than the one before it. That this spending function maintains the familywise error rate is shown in Goodman, Li, and Tiwari (2011).

Example: simulating data and detailed prediction

In order to make event predictions, a sufficiently clean dataset including accrual times and the relevant time-to-event endpoint data up to the CCOD is needed. In addition, if recruitment is not yet completed, a prediction regarding the accrual of the remaining patient is required as well.

In this section, a detailed example of all steps required for an event prediction is provided based on simulated trial data from rpact. In the next section, a simpler minimal analysis with only the key steps is shown.

First, let us load the necessary packages:

# load necessary packages
library(rpact)          # to simulate illustrative dataset
library(eventTrack)     # event tracking methodology
library(fitdistrplus)   # to estimate Weibull fit
library(knitr)          # to prettify some outputs

Simulating an illustrative dataset

We start by illustrating how to predict the number of events of a clinical trial over time at the design stage using rpact:

# ------------------------------------------------------------
# trial details 
# ------------------------------------------------------------
alpha <- 0.05
beta <- 0.2

# design
design <- getDesignGroupSequential(informationRates = 1, typeOfDesign = "asOF", 
                                   sided = 1, alpha = alpha, beta = beta)

# max number of patients
maxNumberOfSubjects1 <- 500
maxNumberOfSubjects2 <- 500
maxNumberOfSubjects <- maxNumberOfSubjects1 + maxNumberOfSubjects2

# survival hazard
piecewiseSurvivalTime <- c(0, 6, 9)
lambda2 <- c(0.025, 0.04, 0.02)

# effect size
hr <- 0.75

# quantities that determine timing of reporting events
dout1 <- 0
dout2 <- 0
douttime <- 12
accrualTime <- 0
accrualIntensity <- 42

# maximal sample size
samplesize <- getSampleSizeSurvival(design = design, piecewiseSurvivalTime = piecewiseSurvivalTime, 
                        lambda2 = lambda2, hazardRatio = hr, dropoutRate1 = dout1, 
                        dropoutRate2 = dout2, dropoutTime = douttime, accrualTime = accrualTime, 
                        accrualIntensity = accrualIntensity,
                        maxNumberOfSubjects = maxNumberOfSubjects)  
nevents <- as.vector(ceiling(samplesize$eventsPerStage))

# expected number of events overall and per group at 1-60 months after trial start
time <- 0:60 
probEvent <- getEventProbabilities(time = time, 
    piecewiseSurvivalTime = piecewiseSurvivalTime, lambda2 = lambda2, 
    hazardRatio = hr,
    dropoutRate1 = dout1, dropoutRate2 = dout2, dropoutTime = douttime,
    accrualTime = accrualTime, accrualIntensity = accrualIntensity, 
    maxNumberOfSubjects = maxNumberOfSubjects)

# extract expected number of events
expEvent <- probEvent$overallEventProbabilities * maxNumberOfSubjects   # overall
expEventInterv <- probEvent$eventProbabilities1 * maxNumberOfSubjects1  # intervention
expEventCtrl <- probEvent$eventProbabilities2 * maxNumberOfSubjects2   # control

# cumulative number of recruited patients over time
nSubjects <- getNumberOfSubjects(time = time, accrualTime = accrualTime, 
                                 accrualIntensity = accrualIntensity, 
                                 maxNumberOfSubjects = maxNumberOfSubjects)

# plot results
plot(time, nSubjects$numberOfSubjects, type = "l", lwd = 2,
     main = "Number of patients and expected number of events over time",
     xlab = "Time from trial start (months)",
     ylab = "Number of patients / events")
lines(time, expEvent, col = "blue", lwd = 2)
lines(time, expEventCtrl, col = "green")
lines(time, expEventInterv, col = "orange")
legend(x = 23, y = 850,
       legend = c("Recruitment", "Expected events - All subjects",
                "Expected events - Control", "Expected events - Intervention"),
       col = c("black", "blue", "green", "orange"), lty = 1, lwd = c(2, 2, 1, 1),
       bty = "n")

# add lines with event prediction pre-trial
fa_pred_pretrial <- exactDatesFromMonths(data.frame(1:length(expEvent) - 1, expEvent), nevents)
segments(fa_pred_pretrial, 0, fa_pred_pretrial, nevents, lwd = 2, lty = 2, col = "blue")
segments(0, nevents, fa_pred_pretrial, nevents, lwd = 2, lty = 2, col = "blue")

For the design, we assume there is no interim analysis, implying that we need 299 events using the specified parameters.

We simplified the specification of the underlying piecewise exponential hazard function somewhat, so that at the time of the CCOD below we have already data on all stretches of the piecewise exponential hazard. Otherwise, it would be impossible to accurately estimate \(S\).

Now assume this trial is running. To generate an illustrative dataset we simulate based on the above assumptions as follows:

seed <- 2020

# ------------------------------------------------------------
# generate a dataset with 1000 patients
# administratively censor after 100 events
# ------------------------------------------------------------

# number of events after which to administratively censor
nevents_admincens <- 100

# generate a dataset after the interim analysis, administratively censor after 100 events
trial1 <- getSimulationSurvival(design,
    piecewiseSurvivalTime = piecewiseSurvivalTime, lambda2 = lambda2,
    hazardRatio = hr, dropoutRate1 = dout1, dropoutRate2 = dout2, dropoutTime = douttime,
    accrualTime = accrualTime, accrualIntensity = accrualIntensity,
    plannedEvents = nevents_admincens, directionUpper = FALSE,
    maxNumberOfSubjects = maxNumberOfSubjects, maxNumberOfIterations = 1,
    maxNumberOfRawDatasetsPerStage = 1, seed = seed)
trial2 <- getRawData(trial1)

# clinical cutoff date for this analysis
ccod <- trial2$lastObservationTime[1]

Assume that after 100 events we clean the dataset and want to predict the timepoint when the final number of 299 events happens. This cleaning milestone is reached after 14 months. The generated dataset looks as follows:

kable(head(trial2, 10)[, 3:ncol(trial2)], row.names = FALSE)

subjectId	accrualTime	treatmentGroup	survivalTime	dropoutTime	lastObservationTime	timeUnderObservation	event	dropoutEvent
1	0.0238095	1	64.9008011	NA	13.99707	13.9732632	FALSE	FALSE
2	0.0476190	2	20.5623950	NA	13.99707	13.9494536	FALSE	FALSE
3	0.0714286	1	59.7432789	NA	13.99707	13.9256441	FALSE	FALSE
4	0.0952381	2	27.8982841	NA	13.99707	13.9018346	FALSE	FALSE
5	0.1190476	1	7.1265000	NA	13.99707	7.1265000	TRUE	FALSE
6	0.1428571	2	2.7904861	NA	13.99707	2.7904861	TRUE	FALSE
7	0.1666667	1	6.8596179	NA	13.99707	6.8596179	TRUE	FALSE
8	0.1904762	2	20.4710395	NA	13.99707	13.8065965	FALSE	FALSE
9	0.2142857	1	0.1379221	NA	13.99707	0.1379221	TRUE	FALSE
10	0.2380952	2	43.9063079	NA	13.99707	13.7589774	FALSE	FALSE

The dataset trial2 contains simulated event times and accrual times for all patients, i.e. also those patients that have been accrued later than 14 months. This means that by the timepoint our event prediction happens in reality these patients would not have been recruited yet. In our dataset, these are 413 patients. In our event prediction, we need to take this into account. The dataset is therefore reorganized as follows:

# ------------------------------------------------------------
# generate a dataset with PFS information:
# - blinded to treatment assignment
# - remove patients recruited *after* clinical cutoff date
# ------------------------------------------------------------
pfsdata <- subset(trial2, subset = (timeUnderObservation > 0), 
                  select = c("subjectId", "timeUnderObservation", "event"))
colnames(pfsdata) <- c("pat", "pfs", "event")
pfsdata[, "event"] <- as.numeric(pfsdata[, "event"])

# define variables with PFS time and censoring indicator
pfstime <- pfsdata[, "pfs"]
pfsevent <- pfsdata[, "event"]

Estimate \(S\)

With this we can now estimate \(S\) based on the patients already accrued and using the hybrid model. First, let us plot the estimated piecewise constant hazard functions together with a kernel estimate of the hazard function of the 587 patients that already have PFS data, either with an event or censored:

# --------------------------------------------------
# analyze hazard functions
# --------------------------------------------------
oldpar <- par(las = 0, mar = c(5.1, 4.1, 4.1, 2.1))
par(las = 1, mar = c(4.5, 5.5, 3, 1))

# kernel estimate of hazard function
h.kernel1 <- muhaz(pfstime, pfsevent)
plot(h.kernel1, xlab = "time (months)", ylab = "", col = grey(0.5), lwd = 2, 
     main = "hazard function", xlim = c(0, max(pfstime)), ylim = c(0, 0.06))
par(las = 3)
mtext("estimates of hazard function", side = 2, line = 4)
rug(pfstime[pfsevent == 1])

# MLE of exponential hazard, i.e. no changepoints
exp_lambda <- sum(pfsevent) / sum(pfstime)
segments(0, exp_lambda, 20, exp_lambda, lwd = 1, col = 2, lty = 1)

# add estimated piecewise exponential hazards
select <- c(1, 3, 5)
legend("topleft", legend = c("kernel estimate of hazard", "exponential hazard", 
                             paste("#changepoints = ", 
     select, sep = ""), "true hazard"), col = c(grey(0.5), 2, 3:(length(select) + 2), 1), 
     lty = c(1, 1, rep(2, length(select)), 1), lwd = c(2, 1, rep(2, length(select)), 2), 
     bty = "n")
for (i in 1:length(select)){
  pe1 <- piecewiseExp_MLE(pfstime, pfsevent, K = select[i])
  tau.i <- c(0, pe1$tau, max(pfstime))
  lambda.i <- c(pe1$lambda, tail(pe1$lambda, 1))  
  lines(tau.i, lambda.i, type = "s", col = i + 2, lwd = 2, lty = 2)
  segments(max(tau.i), tail(lambda.i, 1), 10 * max(pfstime), 
           tail(lambda.i, 1), col = i + 2, lwd = 2)
}

# since we simulated data we can add the truth
lines(stepfun(piecewiseSurvivalTime[-1], lambda2), lty = 1, col = 1, lwd = 2)

As a next step, we perform the sequential testing for detection of a changepoint, starting from an estimated hazard with \(K = 5\) changepoints:

# --------------------------------------------------
# starting from a piecewise Exponential hazard with 
# K = 5 change points, infer the last "significant"
# change point
# --------------------------------------------------
pe5 <- piecewiseExp_MLE(time = pfstime, event = pfsevent, K = 5)
pe5.tab <- piecewiseExp_test_changepoint(peMLE = pe5, alpha = 0.05)
cp.select <- max(c(0, as.numeric(pe5.tab[, "established change point"])), na.rm = TRUE)
kable(subset(pe5.tab, select = c("k", "H_0", "alpha_{k-1}", "p-value", "reject", 
                                 "established change point")), row.names = FALSE)

k	H_0	alpha_{k-1}	p-value	reject
2	lambda1 - lambda2	0.0250000	0.3913751	0
3	lambda2 - lambda3	0.0125000	0.0098730	1
4	lambda3 - lambda4	0.0062500	0.0023362	1
5	lambda4 - lambda5	0.0031250	0.0943560	0
6	lambda5 - lambda6	0.0015625	0.0408903	0

Proceeding in the sequential fashion as described above no changepoint can be established based on this data. This, because we do not reject the first null hypothesis and thus the procedure stops, eventhough the \(p\)-value for \(k = 3, 4\) are smaller than the corresponding local significance levels. The implication of this is that we do not select a changepoint, i.e. no hybrid exponential model, but simply estimate \(S\) based on a constant hazard, i.e. a simple exponential model.

Prediction of the timepoint of the FA

The selected estimator for \(S\) is the one without changepoint. With this estimated survival function we can now proceed and predict the timepoint when the targeted 299 events happen. To this end we will compute predicted events on a monthly grid. For illustration we also plot the estimated piecewise exponential survival functions with \(K = 5\) changepoints and provide the predictions based on that estimate of \(S\).

Note that for the prediction we now also include those patients that will only be accrued after our CCOD, as discussed above.

# ------------------------------------------------------------
# projection for patients to be recruited after CCOD
# 42 patients / month, and we have maxNumberOfSubjects - length(pfstime)
# left to randomize. Assume they show up uniformly distributed
# this accrual is relative to the CCOD
# ------------------------------------------------------------
accrual_after_ccod <- 1:(maxNumberOfSubjects - length(pfstime)) / accrualIntensity 

# Kaplan-Meier and hybrid Exponential tail fits, for different changepoints
par(las = 1, mar = c(4.5, 4.5, 3, 1))
plot(survfit(Surv(pfstime, pfsevent) ~ 1), mark = "", xlim = c(0, 30), 
     ylim = c(0, 1), conf.int = FALSE, xaxs = "i", yaxs = "i",
     main = "estimated survival functions", xlab = "time", 
     ylab = "survival probability", col = grey(0.75), lwd = 5)

# how far out should we predict monthly number of events?
future.units <- 15
tab <- matrix(NA, ncol = 2, nrow = future.units)
ts <- seq(0, 100, by = 0.01)

# the function predictEvents takes as an argument any survival function
# hybrid exponential with cp.select
St1 <- function(t0, time, event, cp){
     return(hybrid_Exponential(t0 = t0, time = time, event = event, 
     changepoint = cp))}

pe1 <- predictEvents(time = pfstime, event = pfsevent, 
          St = function(t0){St1(t0, time = pfstime, 
          event = pfsevent, cp.select)}, accrual_after_ccod, 
          future.units = future.units)
tab[, 1] <- pe1[, 2]
lines(ts, St1(ts, pfstime, pfsevent, cp.select), col = 2, lwd = 2)

# for illustration add piecewise exponential with five changepoints
St2 <- function(t0, time, event, tau, lambda){
      return(1 - getPiecewiseExponentialDistribution(time = t0, 
                        piecewiseSurvivalTime = tau, 
                        piecewiseLambda = lambda))}

pe2 <- predictEvents(time = pfstime, event = pfsevent, 
          St = function(t0){St2(t0, time = pfstime, 
          event = pfsevent, tau = c(0, pe5$tau), 
          lambda = pe5$lambda)}, accrual_after_ccod, 
          future.units = future.units)
tab[, 2] <- pe2[, 2]
lines(ts, St2(ts, pfstime, pfsevent, c(0, pe5$tau), pe5$lambda), col = 3, lwd = 2)

# since we have simulated data we can add the truth
tp <- seq(0, 100, by = 0.01) 
lines(tp, 1 - getPiecewiseExponentialDistribution(time = tp, 
                        piecewiseSurvivalTime = piecewiseSurvivalTime, 
                        piecewiseLambda = lambda2), col = 4, lwd = 2, lty = 1)

# legend
legS <- c("Kaplan-Meier", "hybrid exponential", 
          "piecewise exponential K = 5", "true S")
legend("bottomleft", legS, col = c(grey(0.75), 2:4), 
       lwd = c(5, 2, 2, 2), bty = "n", lty = 1)

# prettify the output
tab <- data.frame(pe1[, 1], round(tab, 1))
colnames(tab) <- c("month", legS[2:3])

Since we can compare our estimates to the truth we see that the hybrid exponential model actually not too badly estimates this truth. It is important to note that based on the generic formula in the methods section above we know that the latest timepoint \(S\) has to be evaluated at is the time of the last censored patient + future.units, which in our case amounts to 29 months, i.e. we have to extrapolate 15 months beyond the last observed PFS time! Looking at the plot of the estimated survival functions we see how variable the various estimates are at 29 months, implying high dependence of the predicted analysis timepoint on the selected model.

In any case, based on these survival functions we then get the following point predictions of event numbers over time:

kable(tab, row.names = FALSE)

month	hybrid exponential	piecewise exponential K = 5
1	113.7	112.8
2	128.1	125.8
3	143.2	139.5
4	159.1	153.8
5	175.7	168.5
6	193.0	183.5
7	210.9	198.8
8	229.5	214.6
9	248.7	230.9
10	268.5	247.6
11	288.2	264.0
12	307.3	279.6
13	326.0	294.4
14	344.1	308.5
15	361.8	322.3

Note the sensitivity of the estimated timepoint on the estimate of \(S\) in the table above.

Our sequential test suggested that we use the model without a changepoint, i.e. a standard exponential fit. This corresponds to the column hybrid exponential. Looking at this column we find that the targeted number of 299 events happens between time unit 11 and 12 after the current CCOD. Linearly interpolating we can get the “precise” timepoint as follows:

fa_pred_ccod <- exactDatesFromMonths(tab[, c("month", "hybrid exponential")], nevents)
fa_pred_ccod

## [1] 11.56545

On the timescale of trial start we need to add the CCOD to get:

fa_pred_ccod + ccod

## [1] 25.56252

Now, let us compare our prediction of when 299 events happen with what we predicted before starting the trial. As we are blinded once the trial started we only re-plot the initial prediction for the entire trial, not separate per arm.

# ------------------------------------------------------------
# plot event predictions pre-trial and after CCOD
# ------------------------------------------------------------
par(las = 1, mar = c(4.5, 4.5, 3, 1))
plot(time, nSubjects$numberOfSubjects, type = "l", lwd = 5, col = grey(0.75), 
     main = "Number of patients and expected number of events over time",
     xlab = "Time from trial start (months)",
     ylab = "Number of patients / events")

# expected events at design stage
lines(time, expEvent, col = 4, lwd = 3, lty = 2)

# patients already recruited
tot_accrual1 <- subset(trial2, select = "accrualTime", subset = timeUnderObservation > 0)[, 1]

# add those that will only be recruited after CCOD
tot_accrual <- c(tot_accrual1, accrual_after_ccod + ccod)
tot_accrual_unit <- cumsum(table(cut(tot_accrual, 0:59, labels = FALSE)))

# actually not much to see, b/c simulated data perfectly uniform
lines(1:length(tot_accrual_unit), tot_accrual_unit, col = 1) 

# now add event prediction based on hybrid exponential and 100 events
lines(tab$month + ccod, tab[, "hybrid exponential"], col = 2, lwd = 3)

# add lines with prediction based on CCOD data
segments(fa_pred_ccod + ccod, 0, fa_pred_ccod + ccod, nevents, lwd = 3, col = 2, lty = 3)
segments(0, nevents, fa_pred_ccod + ccod, nevents, lwd = 3, col = 2, lty = 3)

# finally, we add observed events over time
# note the x-axis of the plot (time from trial start), so we need to add accrual time
events_trial_time <- sort((pfstime + tot_accrual1)[pfsevent == 1])
lines(stepfun(events_trial_time, 0:sum(pfsevent)), pch = NA, col = "orange", lwd = 2)

# legend
legend(x = 20, y = 930, c("Planned recruitment at design stage", 
                          "Recruitment (up to CCOD + prediction)", 
                          "Expected events at design stage", 
                          "Observed events", 
                          "Expected events (after CCOD prediction)"),
       col = c(grey(0.75), 1, 4, "orange", 2), lty = c(1, 1, 2, 1, 1), 
       lwd = c(3, 1, 3, 2, 3), bty = "n")

# revert back to initial plot parameters
par(oldpar)

So, before the trial started, based on our initial assumptions, we predicted that 299 events would happen after 28.6 months. Our updated prediction after 100 events is now 25.6 months, using the hybrid exponential approach. This prediction is actually quite accurate, taking into account that it relies on only 100 patients with event out of 1000 patients. We have also added the actually observed events over time. As we simulate from the truth the observation is very close to the expectation - this might not be the case in an actual trial.

In general, event predictions come with high variability, see also the strategic discussion below. Note that we have (at least) two sources of variability:

Accrual of the 413 patients still to be accrued is only a projection.
Variability in estimation of \(S\).

Using bootstrap we can easily quantify the uncertainty in estimation of \(S\), for a fixed accrual projection:

# ------------------------------------------------------------
# compute a confidence interval for our event prediction using bootstrap
# ------------------------------------------------------------

# do not run as takes a few minutes, code only here for illustration
if (1 == 0){
boot1 <- bootstrapTimeToEvent(time = pfstime, event = pfsevent, 
          interim.gates = nevents, future.units = 20, n = length(pfstime), 
          M0 = 1000, K0 = 5, alpha = 0.05, accrual = accrual_after_ccod, seed = 2020)
ci <- quantile(boot1$event.dates[, 1], c(0.025, 0.975))
}

# hard-code result and add ccod
ci <- c(10.26989, 12.50578) + ccod
ci

## [1] 24.26696 26.50285

So a 95% confidence interval around our predicted CCOD for the final analysis, 25.6 months, ranges from 24.3 to 26.5 months. Note that we implement the entire model selection algorithm into the bootstrap, i.e. the changepoint is determined individually for every bootstrap sample.

Using parametric Weibull instead of hybrid exponential estimate

If instead the hybrid exponential model we would like to use, e.g., a parametric model such as Weibull, we can proceed as follows:

# ------------------------------------------------------------
# event prediction based on pure Weibull assumption
# ------------------------------------------------------------

# prepare dataset to fit function "fitdistcens"
pfsdata2 <- data.frame(cbind("left" = pfstime, "right" = pfstime))
pfsdata2[pfsevent == 0, "right"] <- NA

# estimate parameters based on Weibull distribution
fit_weibull <- suppressWarnings(fitdistcens(pfsdata2, "weibull"))

# event prediction based on this estimated survival function
# simply specify survival function based on above fit
pred_weibull <- predictEvents(time = pfstime, event = pfsevent, 
                      St = function(t0){1 - pweibull(t0, 
                                                     shape = fit_weibull$estimate["shape"], 
                                                     scale = fit_weibull$estimate["scale"])}, 
                      accrual = accrual_after_ccod, future.units = future.units)
ccod_weibull <- exactDatesFromMonths(pred_weibull, nevents)
ccod_weibull

## [1] 12.27657

ccod_weibull + ccod

## [1] 26.27364

So the predicted CCOD for the FA based on a Weibull fit amounts to 26.3, very similar to what we receive based on the hybrid exponential model.

Example: minimal analysis

This section collects the code that is minimally needed to get to a timepoint of a prediction, without the above more detailed discussion.

# ------------------------------------------------------------
# input data (e.g. OS or PFS)
# we use the data simulated above
# alternatively, simply use
# pfsdata <- read.csv("C:/my path/pfs.csv")
# --------------------------------------------------

# accrual after ccod
accrual_after_ccod <- 1:(maxNumberOfSubjects - length(pfstime)) / accrualIntensity 

# infer changepoint
pe5 <- piecewiseExp_MLE(time = pfstime, event = pfsevent, K = 5)
pe5.tab <- piecewiseExp_test_changepoint(peMLE = pe5, alpha = 0.05)
cp.selected <- suppressWarnings(max(as.numeric(pe5.tab[, "established change point"]), 
                                             na.rm = TRUE))
cp.selected <- max(cp.selected, 0)

# compute predictions for selected hybrid exponential model, i.e. chosen changepoint
future.units <- 15
St1 <- function(t0, time, event, cp){
     return(hybrid_Exponential(t0 = t0, time = time, event = event, 
     changepoint = cp))}

pe1 <- predictEvents(time = pfstime, event = pfsevent, 
          St = function(t0){St1(t0, time = pfstime, 
          event = pfsevent, cp.selected)}, accrual_after_ccod, 
          future.units = future.units)

# find exact timepoint when targeted number of events are reached through linear interpolation
# add computation of confidence interval using code from just above
fa_pred_ccod <- exactDatesFromMonths(tab[, c("month", "hybrid exponential")], nevents)
fa_pred_ccod + ccod

## [1] 25.56252

Possible extensions

The method implemented in eventTrack currently does not account for drop-out.

The formula for \(m(t, S)\) introduced in the methodology section above is generic, i.e. remains valid for event predictions. What potentially can be improved in a given example is

the quality of the accrual projection and
the extrapolation of the estimate of \(S\) beyond where we have data.

The hybrid exponential model is a compromise between a fully parametric and a fully nonparametric estimation method. However, there might potentially be better methods around for extrapolation of survival curves. Such extrapolation is the bread-and-butter of Health Economists and I suspect they could offer good approaches, e.g. using external data to inform extrapolation (Jackson et al. (2017), Guyot et al. (2017)) of Bayesian model averaging. However, the effort put in any modelling effort has to be balanced against a potential gain in precision in estimates. Typically, the approach described above provides “good enough” predictions in many instances.

What is described in this document applies to one treatment arm, or equivalently, to prediction blinded to treatment assignment. If you are unblinded to the treatment assignment and want to use that information for prediction, then you can perform a separate prediction of events per arm (potentially also using an estimation method for \(S\) that is different between the arms) and then add these up.

Strategic context

Implementing an event prediction based on data and getting the relevant numbers, point prediction + confidence interval, out is certainly an important aspect. However, in a drug development context this is only the starting point. Strategic considerations are important. Drawing from my experience I list a few aspects that I found important in the past:

Only do the above exercise on sufficiently clean data. Predictions are worthless if many events are still missing because of a cleaning lag.
This implies that I would strongly recommend not to update projections on a regular basis, even if teams ask for it. Stick with the pre-trial projections until you have a clean dataset for the first time, e.g. at a futility interim analysis. Then only update projections of dates of further IAs and the FA at such milestones.
Communication is key! Explain to teams that these point predictions come with substantial uncertainty. Often, confidence intervals may be much wider than in the above example. Make sure the team understands that this uncertainty is inherent to the problem! This, because of the need for extrapolation + accrual projections if the trial has not finished accrual yet. If a prediction of an analysis milestone, e.g. the FA, has to be substantially shifted based on updated data this does not necessarily point to an earlier mistake by the team or the statistician, but may simply be due to variability. See my BBS talk for an example.
If a trial is approaching a CCOD and, say, only a few events are still missing other approaches might be more meaningful to finally determine the CCOD, such as case-by-case analysis of all patients that are still censored. The moving average approach described by the oncology working group (see details below) might also be useful in that scenario.
Typically, when doing prediction for PFS, investigator-PFS is more up-to-date compared to IRC-PFS. If the latter is your primary endpoint it might still be useful to do event prediction on Inv-PFS and then exploit the “concordance” between the two PFS variants to predict a timepoint when the necessary number of IRC-PFS events will be reached. A very simple example could be: at the time when you have 100 Inv-PFS events you observe (before full IRC read-out, i.e. on the raw data) 80 IRC-PFS events. So if you need a prediction when 299 IRC-PFS events happen then you could take the timepoint when you predict 100 / 80 * 299 = 374 Inv-PFS events.

Other approaches

gestate

The R package gestate, Bell (2020). The vignette describes its functions in a similar fashion as above, and also allows to compute intervals around the prediction. As I understand it, gestate fits several parametric models and then selects the one that “bests” fits the data. For our data this is Weibull, so that we get the same prediction as for the Weibull model above. gestate supplies confidence intervals for the number of events at a given timepoint, but not for the timepoint of the FA CCOD.

It is worth noting that - in contrast to eventTrack - gestate can handle drop-out.

library(gestate)

# accrual: observed + predicted (observed accrual not needed when using eventTrack!)
tot_accrual1 <- subset(trial2, select = "accrualTime", subset = timeUnderObservation > 0)[, 1]
tot_accrual <- c(tot_accrual1, accrual_after_ccod + ccod)
tot_accrual_unit <- cumsum(table(cut(tot_accrual, 0:59, labels = FALSE)))

# Create an RCurve incorporating both observed and predicted recruitment 
recruit <- PieceR(matrix(c(rep(1, length(tot_accrual_unit)), 
                           c(tot_accrual_unit[1], diff(tot_accrual_unit))), ncol = 2), 1)

# do the prediction: need to supply CCOD (ccod), number of events at CCOD (nevents_admincens), 
# and numbers at risk then (length(tot_accrual1))
predictions <- event_prediction(data = pfsdata, Time = "pfs", Event = "event", 
                                censoringOne = FALSE, rcurve = recruit, max_time = 60, 
                                cond_Events = nevents_admincens, 
                                cond_NatRisk = length(tot_accrual1), 
                                cond_Time = ceiling(ccod), units = "Months")

# exact date (not exactly as Weibull fit above, as cond_Time argument needs to be an integer)
pred_gestate <- exactDatesFromMonths(predictions$Summary[, c("Time", "Predicted_Events")], nevents)
pred_gestate

## [1] 26.32032

Info

R version and packages used to generate this report:

R version: R version 4.2.3 (2023-03-15 ucrt)

Base packages: stats / graphics / grDevices / utils / datasets / methods / base

Other packages: gestate / knitr / fitdistrplus / MASS / rpact / eventTrack / muhaz / survival

This document was generated on 2023-08-03 at 12:14:22.

References

Bell, James. 2020. Gestate: Generalised Survival Trial Assessment Tool Environment. https://CRAN.R-project.org/package=gestate.

Fang, L., and Z. Su. 2011. “A hybrid approach to predicting events in clinical trials with time-to-event outcomes.” Contemp Clin Trials 32 (5): 755–59.

Goodman, M. S., Y. Li, and R. C. Tiwari. 2011. “Detecting multiple change points in piecewise constant hazard functions.” J Appl Stat 38 (11): 2523–32.

Guyot, Patricia, Anthony E Ades, Matthew Beasley, Béranger Lueza, Jean-Pierre Pignon, and Nicky J Welton. 2017. “Extrapolation of Survival Curves from Cancer Trials Using External Information.” Medical Decision Making : An International Journal of the Society for Medical Decision Making 37 (May): 353–66. https://doi.org/10.1177/0272989X16670604.

Jackson, Christopher, John Stevens, Shijie Ren, Nick Latimer, Laura Bojke, Andrea Manca, and Linda Sharples. 2017. “Extrapolating Survival from Randomized Trials Using External Data: A Review of Methods.” Medical Decision Making : An International Journal of the Society for Medical Decision Making 37 (May): 377–90. https://doi.org/10.1177/0272989X16639900.

Rufibach, Kaspar. 2016. “Event Projection: Quantify Uncertainty and Manage Expectations of Broader Teams.” Talk at BBS seminar 18th April 2016. https://baselbiometrics.github.io/home/docs/talks/20160428/1_Rufibach.pdf.

———. 2021. eventTrack: Implements Functions and Datasets for Event Tracking.

Event tracking: predict analysis timepoints for ongoing clinical trials

Kaspar Rufibach