In Roche/crmPack: Object-Oriented Implementation of CRM Designs

Objective

When defining a simulation design, we would like to have the option to assess if Stopping rules are met for specific doses of the planned dose grid and not just for the next best dose.

Currently all the stopping rules are assessed at the next best dose and not at a predefined dose of the planned dose grid.

Idea

We can define a wrapper class say StopSpecificDose which takes as slots any stopping rule as well as one specific dose, and then always evaluates this stopping rule at this specific dose (instead of the next best dose).

Maybe we should name it StoppingSpecificDose?

Prototype

Validation function

v_stop_specific_dose <- function(object) {
  v <- Validate()
  v$check(
    test_number(object@dose, finite = TRUE),
    "dose needs to be a single finite number"
  )
  v$result()
}

Rule class

This is the class definition, returning the default constructor:

.StopSpecificDose <- setClass(
  Class = "StopSpecificDose",
  slots = c(
    rule = "Stopping",
    dose = "numeric"
  ),
  contains = "Stopping",
  validity = v_stop_specific_dose
)

Let's also define a constructor function such that it can be easily used.

StopSpecificDose <- function(rule, dose) {
  .StopSpecificDose(rule = rule, dose = dose)
}

Rule method

Here comes the most interesting part. Note that here dose is the next best dose.

setMethod("stopTrial",
  signature =
    signature(
      stopping = "StopSpecificDose",
      dose = "numeric",
      samples = "ANY",
      model = "ANY",
      data = "Data"
    ),
  def =
    function(stopping, dose, samples, model, data, ...) {
      # Make sure that the specific dose is part of the dose grid.
      assert_subset(x = stopping@dose, choices = data@doseGrid)

      # Now we evaluate the original (wrapped) stopping rule at the specific dose.
      result <- stopTrial(
        stopping = stopping@rule,
        dose = stopping@dose,
        samples = samples,
        model = model,
        data = data,
        ...
      )

      # We can now try to correct the text from the original stopping rule.
      original_text <- attr(result, "message")
      new_text <- gsub(pattern = "next best", replacement = "specific", x = original_text)
      attr(result, "message") <- new_text

      result
    }
)

Note that in production later we could consider just removing the "next best" part from existing stopping rule messages. Or provide some wrapper text around original text or similar.

Trying it out

For example we could evaluate one rule at the highest dose.

highest_dose_safe <- StopSpecificDose(
  rule = StoppingTargetProb(target=c(0, 0.3), prob=0.8),
  dose = 80
)
max_patients <- StoppingMinPatients(nPatients = 20)
patients_near_dose <- StoppingPatientsNearDose(nPatients = 3, percentage = 0)
my_stopping <- highest_dose_safe | max_patients | patients_near_dose

Then we create some data and model with MCMC samples.

data <- Data(
  x = c(0.1, 0.5, 1.5, 3, 6, 10, 10, 10),
  y = c(0, 0, 0, 0, 0, 0, 1, 0),
  cohort = c(0, 1, 2, 3, 4, 5, 5, 5),
  doseGrid =
    c(
      0.1, 0.5, 1.5, 3, 6,
      seq(from = 10, to = 80, by = 2)
    )
)

model <- LogisticLogNormal(
  mean = c(-0.85, 1),
  cov =
    matrix(c(1, -0.5, -0.5, 1),
      nrow = 2
    ),
  ref_dose = 50
)

options <- McmcOptions(
  burnin = 100,
  step = 2,
  samples = 2000
)
set.seed(94)
samples <- mcmc(data, model, options)

next_best <- NextBestNCRM(
  target = c(0.2, 0.35),
  overdose = c(0.35, 1),
  max_overdose_prob = 0.25
)

doseRecommendation <- nextBest(next_best,
  doselimit = 100,
  samples = samples, model = model, data = data
)
doseRecommendation$value

20 is the next best dose.

Now we can evaluate the stopping rule:

stopTrial(
  stopping = my_stopping, 
  dose = doseRecommendation$value,
  samples = samples, 
  model = model, 
  data = data
)

This shows that it works, because the text for the first stopping rule is citing the dose 80 (which we specified above as the specific dose to evaluate the StopSpecificDose rule), while the last one cites the next best dose 20.

Roche/crmPack documentation built on June 30, 2024, 1:31 a.m.