TITE-CRM and TITE-NBG - Time-to-Event Designs

Source: vignettes/A280-TITE.Rmd

Introduction

TITE-CRM

Y. K. Cheung and Chappell (2000) introduced TITE-CRM as a variant of the regular CRM (O’Quigley, Pepe, and Fisher 1990) that handles late-onset toxicities. Dose-finding trials tend to use a short toxicity window after the commencement of therapy, during which each patient is evaluated for the presence or absence of dose-limiting toxicity (DLT). This approach works well in treatments like chemotherapy where toxic reactions are expected to manifest relatively quickly. In contrast, one of the hallmarks of radiotherapy, for instance, is that related adverse reactions can manifest many months after the start of treatment. A similar phenomenon may arise with immunotherapies.

In adaptive dose-finding clinical trials, where doses are selected mid-trial in response to the outcomes experienced by patients evaluated hitherto, late-onset toxic events present a distinct methodological challenge. Naturally, the toxicity window will need to be long enough to give the trial a good chance of observing events of interest. If, however, we wait until each patient completes the evaluation window before using their outcome to forecast the best dose, the trial may take an infeasibly long time and ignore pertinent interim data.

TITE-CRM presents a solution by introducing the notion of a partial tolerance event. If a patient is half way through the evaluation window and has not yet experienced toxicity, we may say that they have experienced half a tolerance. This simple novelty allows partial information to be used in dose-recommendation decisions. If the patient goes on to complete the window with no toxic reaction, they will be regarded as having completely tolerated treatment, as is normally the case with CRM and other dose-finding algorithms. This notion of partial events is only applied to tolerances, however. If a patient experiences toxicity at any point during the evaluation window, they are immediately regarded as having experienced 100% of a DLT event.

To illustrate TITE-CRM mathematically, we start with the likelihood from the plain vanilla CRM. Let YiY_i be a random variable taking values {0,1}\{0, 1\} reflecting the absence and presence of DLT respectively in patient ii. A patient administered dose xix_i has estimated probability of toxicity F(xi,θ)F(x_i, \theta), where θ\theta represents the set of model parameters. The likelihood component arising from patient ii is

F(xi,θ)Yi(1−F(xi,θ))1−Yi F(x_i, \theta)^{Y_i} (1 - F(x_i, \theta))^{1-Y_i}

and the aggregate likelihood after the evaluation of JJ patients is

LJ(θ)=∏i=1J{F(xi,θ)}Yi{1−F(xi,θ)}1−Yi L_J(\theta) = \prod_{i=1}^J \left\{ F(x_i, \theta) \right\}^{Y_i} \left\{ 1 - F(x_i, \theta) \right\}^{1-Y_i}

Y. K. Cheung and Chappell (2000) observed that each patient may provide a weight, wiw_i, reflecting the extend to which their outcome has been evaluated. The weighted likelihood is

LJ(θ)=∏i=1J{wiF(xi,θ)}Yi{1−wiF(xi,θ)}1−Yi L_J(\theta) = \prod_{i=1}^J \left\{ w_i F(x_i, \theta) \right\}^{Y_i} \left\{ 1 - w_i F(x_i, \theta) \right\}^{1-Y_i}

TITE-CRM weights the outcomes according to the extend to which patients have completed the evaluation period. To illustrate the design, we reproduce the example given on p.124 of Ying Kuen Cheung (2011). Four patients have been treated at dose-level 3 and all are part-way through the 126-day toxicity evaluation window.

The authors use the empiric model so that there is one parameter, θ=β\theta = \beta, the dose-toxicity relation is F(xi,β)=xiexp(β)F(x_i, \beta) = x_i^{exp(\beta)}, and a N(0,σβ2)N(0, \sigma_{\beta}^2) prior is specified on β\beta.

TITE-NBG

Neuenschwander, Branson, and Gsponer (2008) (NBG) introduced a derivative of the CRM for dose-escalation clinical trials using a two-parameter model (see the NBG vignette).

The authors did not introduce a time-to-event variant of their design but it was simple to add one to escalation using the same method presented above with a weight parameter, ww. We demonstrate that method here alongside TITE-CRM.

Implementation in escalation

TITE-CRM

As with the regular CRM, we require a dose-toxicity skeleton and a target toxicity value. For illustration, we use the same parameters as used in the plain CRM vignette:

library(escalation)

skeleton <- c(0.05, 0.12, 0.25, 0.40, 0.55)
target <- 0.25

a0 <- 3
beta_sd <- sqrt(1.34)

a0 is the fixed intercept value and beta_sd is the SD of the slope parameter. We then have design:

model <- get_dfcrm_tite(
  skeleton = skeleton, 
  target = target,
  model = "logistic", 
  intcpt = a0, 
  scale = beta_sd
)

Elsewhere in escalation, we have represented outcomes using character strings. The complexity of specifying patient-level weights prevents that approach in time-to-event trials. Instead we represent outcomes in data-frames. For example:

outcomes <- data.frame(
  dose = c(1, 1, 2, 2, 3, 3),
  tox = c(0, 0, 0, 0, 1, 0),
  weight = c(1, 1, 1, 0.9, 1, 0.5),
  cohort = c(1, 2, 3, 4, 5, 6)
)

outcomes
#>   dose tox weight cohort
#> 1    1   0    1.0      1
#> 2    1   0    1.0      2
#> 3    2   0    1.0      3
#> 4    2   0    0.9      4
#> 5    3   1    1.0      5
#> 6    3   0    0.5      6

represents a scenario where we have sequentially had two patients each at dose-levels 1, 2 and 3. The first three patients have been fully evaluated as tolerating treatment (their tox parameters are 0 and their weight parameters are 1). The fifth patient is also fully-weighted because they unfortunately experienced toxicity (their tox parameter is 1 and their weight parameter is also 1). Finally, the fourth and sixth patients are still being evaluated without having experienced toxicity so far (their tox parameters are 0 and their weight parameters are less than 1).

To fit the model to these outcomes, we run:

x <- model %>% fit(outcomes)

The usual generic functions apply:

print(x)
#> Patient-level data:
#> # A tibble: 6 × 5
#>   Patient Cohort  Dose   Tox Weight
#>     <int>  <int> <int> <int>  <dbl>
#> 1       1      1     1     0    1  
#> 2       2      2     1     0    1  
#> 3       3      3     2     0    1  
#> 4       4      4     2     0    0.9
#> 5       5      5     3     1    1  
#> 6       6      6     3     0    0.5
#> 
#> Dose-level data:
#> Warning: `...` must be empty in `format.tbl()`
#> Caused by error in `format_tbl()`:
#> ! `...` must be empty.
#> ✖ Problematic argument:
#> • digits = 3
#> # A tibble: 6 × 9
#>   dose     tox     n empiric_tox_rate mean_prob_tox median_prob_tox admissible
#>   <ord>  <dbl> <dbl>            <dbl>         <dbl>           <dbl> <lgl>     
#> 1 NoDose     0     0              0          0               0      TRUE      
#> 2 1          0     2              0          0.0748          0.0748 TRUE      
#> 3 2          0     2              0          0.164           0.164  TRUE      
#> 4 3          1     2              0.5        0.310           0.310  TRUE      
#> 5 4          0     0            NaN          0.460           0.460  TRUE      
#> 6 5          0     0            NaN          0.599           0.599  TRUE      
#> # ℹ 2 more variables: recommended <lgl>, Skeleton <dbl>
#> 
#> The model targets a toxicity level of 0.25.
#> The model advocates continuing at dose 3.

Note the Weight columns in the patient-level output.

recommended_dose(x)
#> [1] 3

TITE-NBG

Let us reuse the design presented and justified by Neuenschwander, Branson, and Gsponer (2008) and demonstrated in the NBG vignette:

dose <- c(1, 2.5, 5, 10, 15, 20, 25, 30, 40, 50, 75, 100, 150, 200, 250)

model <- get_trialr_nbg_tite(
  real_doses = dose, d_star = 250, target = 0.3,
  alpha_mean = 2.15, alpha_sd = 0.84,
  beta_mean = 0.52, beta_sd = 0.8,
  seed = 2020
)

To fit the model to the outcomes from the TITE-CRM example, we run:

x <- model %>% fit(outcomes)

The usual generic functions apply:

print(x)
#> Patient-level data:
#> # A tibble: 6 × 5
#>   Patient Cohort  Dose   Tox    Weight
#>     <int>  <int> <int> <int> <dbl[1d]>
#> 1       1      1     1     0       1  
#> 2       2      2     1     0       1  
#> 3       3      3     2     0       1  
#> 4       4      4     2     0       0.9
#> 5       5      5     3     1       1  
#> 6       6      6     3     0       0.5
#> 
#> Dose-level data:
#> Warning: `...` must be empty in `format.tbl()`
#> Caused by error in `format_tbl()`:
#> ! `...` must be empty.
#> ✖ Problematic argument:
#> • digits = 3
#> # A tibble: 16 × 9
#>    dose     tox     n empiric_tox_rate mean_prob_tox median_prob_tox admissible
#>    <ord>  <dbl> <dbl>            <dbl>         <dbl>           <dbl> <lgl>     
#>  1 NoDose     0     0              0          0               0      TRUE      
#>  2 1          0     2              0          0.0976          0.0569 TRUE      
#>  3 2          0     2              0          0.163           0.124  TRUE      
#>  4 3          1     2              0.5        0.239           0.210  TRUE      
#>  5 4          0     0            NaN          0.343           0.330  TRUE      
#>  6 5          0     0            NaN          0.416           0.417  TRUE      
#>  7 6          0     0            NaN          0.472           0.481  TRUE      
#>  8 7          0     0            NaN          0.518           0.532  TRUE      
#>  9 8          0     0            NaN          0.555           0.573  TRUE      
#> 10 9          0     0            NaN          0.614           0.637  TRUE      
#> 11 10         0     0            NaN          0.659           0.683  TRUE      
#> 12 11         0     0            NaN          0.733           0.758  TRUE      
#> 13 12         0     0            NaN          0.780           0.806  TRUE      
#> 14 13         0     0            NaN          0.835           0.859  TRUE      
#> 15 14         0     0            NaN          0.866           0.889  TRUE      
#> 16 15         0     0            NaN          0.886           0.909  TRUE      
#> # ℹ 2 more variables: recommended <lgl>, RealDose <dbl>
#> 
#> The model targets a toxicity level of 0.3.
#> The model advocates continuing at dose 4.
recommended_dose(x)
#> [1] 4

Dose paths

Dose-paths do not make sense for time-to-event designs because there are practically infinite trial states once continuous patient-level weights are considered. Dose-paths are not implemented.

Simulation 

true_prob_tox <- c(0.25, 0.35, 0.5, 0.6, 0.7, 0.8)

For the sake of speed, we will run just a few iterations:

num_sims <- 20

In real life, however, we would naturally run many iterations.

Let us restrict the design to a sample size of 12 for a quick illustration. Running the simulation:

model <- get_dfcrm_tite(
  skeleton = skeleton, 
  target = target,
  model = "logistic", 
  intcpt = a0, 
  scale = beta_sd
) %>% 
  stop_at_n(n = 12)

set.seed(2025)
sims <- model %>% 
  simulate_trials(
    num_sims = num_sims, 
    true_prob_tox = true_prob_tox,
    max_time = 10
  )

we see that from this small sample size that the low doses are most likely to be recommended:

prob_recommend(sims)
#> NoDose      1      2      3      4      5 
#>   0.00   0.75   0.20   0.05   0.00   0.00

with most patients treated at low dose-levels too:

colMeans(n_at_dose(sims))
#>    1    2    3    4    5 
#> 9.70 0.15 1.40 0.55 0.20

The simulated trial durations could be of interest in a time-to-event trial:

trial_duration(sims)
#>  [1] 19.64932 17.92124 20.81784 20.33176 19.79155 19.20623 24.85348 23.14870
#>  [9] 17.31764 23.62735 23.21265 18.13445 23.99437 21.26878 20.89885 20.28926
#> [17] 24.72482 28.22818 23.73679 26.39005

The max_time parameter was specified so that the design could calculate the patient-level weights (i.e. how much of the observation window had been completed). In the above example, patients were assumed to arrive one at a time with the intra-patient arrival times distributed by an Exponential(1) distribution; this is the default. To override that, we specify a function in the sample_patient_arrivals parameter. The function must take the patient-level data-frame of the prevailing trial data as a parameter (i.e. with columns cohort, patient, dose, tox, and time), and return a data-frame with column time_delta and nn rows containing the arrival deltas of the next nn patients. For example, consider that you want to evaluate patients in cohorts of two. This is somewhat strange because a time-to-event design frees you from the obligation to use cohorts, but it is simple to implement. We could run:

set.seed(2025)
sims <- model %>% 
  simulate_trials(
    num_sims = num_sims, 
    true_prob_tox = true_prob_tox,
    max_time = 10,
    sample_patient_arrivals = function(df) {
      cohorts_of_n(n = 2, mean_time_delta = 1)
    },
    return_all_fits = TRUE
  )

Note: the return_all_fits = TRUE param means that every model fit (all interims and final fits) are returned. We use it here to peak into the way simulations occur. However, in a production run with a large number of iterates and many patients, doing this could lead to memory problems. By default, return_all_fits is FALSE.

The patient arrival times for the (for instance) third simulated trial iterate are random and always increasing:

library(purrr)
#> 
#> Attaching package: 'purrr'
#> The following object is masked from 'package:magrittr':
#> 
#>     set_names
map_dbl(sims$fits[[3]], "time")
#> [1]  0.000000  1.158482  2.835637  4.883962  7.830599  9.916354 10.534122
#> [8] 20.534122

However, the doses given (patients in columns, simulated iterates in rows) appear in pairs to reflect that patients were treated in cohorts of two:

doses_given(sims)
#>       [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12]
#>  [1,]    1    1    3    3    4    4    5    5    3     3     3     3
#>  [2,]    1    1    3    3    4    4    1    1    1     1     1     1
#>  [3,]    1    1    3    3    1    1    1    1    1     1     1     1
#>  [4,]    1    1    3    3    4    4    1    1    1     1     1     1
#>  [5,]    1    1    1    1    1    1    1    1    1     1     1     1
#>  [6,]    1    1    3    3    4    4    1    1    1     1     1     1
#>  [7,]    1    1    1    1    1    1    1    1    1     1     1     1
#>  [8,]    1    1    3    3    1    1    1    1    1     1     1     1
#>  [9,]    1    1    1    1    1    1    1    1    1     1     1     1
#> [10,]    1    1    3    3    1    1    1    1    1     1     1     1
#> [11,]    1    1    3    3    1    1    1    1    1     1     1     1
#> [12,]    1    1    3    3    1    1    1    1    1     1     1     1
#> [13,]    1    1    1    1    1    1    1    1    1     1     1     1
#> [14,]    1    1    3    3    4    4    1    1    1     1     1     1
#> [15,]    1    1    3    3    1    1    1    1    1     1     1     1
#> [16,]    1    1    1    1    1    1    1    1    1     1     1     1
#> [17,]    1    1    3    3    1    1    1    1    1     1     1     1
#> [18,]    1    1    4    4    1    1    1    1    1     1     1     1
#> [19,]    1    1    3    3    1    1    1    1    1     1     1     1
#> [20,]    1    1    3    3    1    1    1    1    1     1     1     1

The inclusion of df as a parameter in the call to sample_patient_arrivals lets you tailor the patient arrival process in many creative ways, including variable cohort sizes, minimum intra-patient gaps, etc.

For more information on running dose-finding simulations, refer to the simulation vignette.

References

Cheung, Y K, and R Chappell. 2000. “Sequential Designs for Phase I Clinical Trials with Late-Onset Toxicities.” Biometrics 56 (4): 1177–82.
Cheung, Ying Kuen. 2011. Dose Finding by the Continual Reassessment Method. New York: Chapman & Hall / CRC Press.
Neuenschwander, Beat, Michael Branson, and Thomas Gsponer. 2008. “Critical aspects of the Bayesian approach to phase I cancer trials.” Statistics in Medicine 27: 2420–39. https://doi.org/10.1002/sim.3230.
O’Quigley, J, M Pepe, and L Fisher. 1990. “Continual Reassessment Method: A Practical Design for Phase 1 Clinical Trials in Cancer.” Biometrics 46 (1): 33–48. https://doi.org/10.2307/2531628.