Creating a Basic ADRS

Introduction

This article describes creating an ADRS ADaM with common oncology endpoint parameters based on RECIST v1.1. Therefore response values are expected as either CR, PR, SD, NON-CR/NON-PD, PD or NE.

Please note that this vignette describes the endpoints which were considered by the admiralonco team as the most common ones. The admiralonco functions used to derive these endpoints provide a certain flexibility, e.g., specifying the reference date or time windows for confirmation or stable disease. If different endpoints or more flexibility is required please read Creating ADRS (Including Non-standard Endpoints).

Examples are currently presented and tested using ADSL (ADaM) and RS, TU (SDTM) inputs. However, other domains could be used. The functions and workflow could similarly be used to create an intermediary ADEVENT ADaM.

Note: All examples assume CDISC SDTM and/or ADaM format as input unless otherwise specified.

Programming Workflow

Read in Data
Pre-processing of Input Records
Derive Progressive Disease Parameter
Derive Response Parameter
Derive Clinical Benefit Parameter
Derive Best Overall Response Parameter
Derive Best Overall Response of CR/PR Parameter
Derive Response Parameters requiring Confirmation
Derive Parameters using Independent Review Facility (IRF)/ Blinded Independent Central Review (BICR) responses
Derive Death Parameter
Derive Last Disease Assessment Parameters
Derive Measurable Disease at Baseline Parameter
Derive AVAL for New Parameters
Assign ASEQ
Add ADSL variables

Read in Data

To start, all data frames needed for the creation of ADRS should be read into the environment. This will be a company specific process. Some of the data frames needed may be ADSL, RS and TU.

For example purpose, the SDTM and ADaM datasets (based on CDISC Pilot test data)—which are included in pharmaversesdtm and pharmaverseadam—are used.

library(admiral)
library(admiralonco)
library(dplyr)
library(pharmaversesdtm)
library(pharmaverseadam)
library(lubridate)
library(stringr)
data("adsl")
data("rs_onco_recist")
data("tu_onco_recist")

rs <- rs_onco_recist
tu <- tu_onco_recist

rs <- convert_blanks_to_na(rs)
tu <- convert_blanks_to_na(tu)

At this step, it may be useful to join ADSL to your RS domain. Only the ADSL variables used for derivations are selected at this step. The rest of the relevant ADSL would be added later.

adsl_vars <- exprs(RANDDT)
adrs <- derive_vars_merged(
  rs,
  dataset_add = adsl,
  new_vars = adsl_vars,
  by_vars = get_admiral_option("subject_keys")
)

Pre-processing of Input Records

The first step involves company-specific pre-processing of records for the required input to the downstream parameter functions. Note that this could be needed multiple times (e.g. once for investigator and once for Independent Review Facility (IRF)/Blinded Independent Central Review (BICR) records). It could even involve merging input data from other sources besides RS, such as ADTR.

This step would include any required selection/derivation of ADT or applying any necessary partial date imputations, updating AVAL (e.g. this should be ordered from best to worst response), and setting analysis flag ANL01FL. Common options for ANL01FL would be to set null for invalid assessments or those occurring after new anti-cancer therapy, or to only flag assessments on or after after date of first treatment/randomization, or rules to cover the case when a patient has multiple observations per visit (e.g. by selecting worst value). Another consideration could be extra potential protocol-specific sources of Progressive Disease such as radiological assessments, which could be pre-processed here to create a PD record to feed downstream derivations.

For the derivation of the parameters it is expected that the subject identifier variables (usually STUDYID and USUBJID) and ADT are a unique key. This can be achieved by deriving an analysis flag (ANLzzFL). See Derive ANL01FL for an example.

The below shows an example of a possible company-specific implementation of this step.

Select Overall Response Records and Set Parameter Details

In this case we use the overall response records from RS from the investigator as our starting point. The parameter details such as PARAMCD, PARAM etc will always be company-specific, but an example is shown below so that you can trace through how these records feed into the other parameter derivations.

adrs <- adrs %>%
  filter(RSEVAL == "INVESTIGATOR" & RSTESTCD == "OVRLRESP") %>%
  mutate(
    PARAMCD = "OVR",
    PARAM = "Overall Response by Investigator",
    PARCAT1 = "Tumor Response",
    PARCAT2 = "Investigator",
    PARCAT3 = "RECIST 1.1"
  )

Partial Date Imputation and Deriving `ADT`, `ADTF`, `AVISIT` etc

If your data collection allows for partial dates, you could apply a company-specific imputation rule at this stage when deriving ADT. For this example, here we impute missing day to last possible date.

adrs <- adrs %>%
  derive_vars_dt(
    dtc = RSDTC,
    new_vars_prefix = "A",
    highest_imputation = "D",
    date_imputation = "last"
  ) %>%
  mutate(AVISIT = VISIT)

Derive `AVALC` and `AVAL`

Here we populate AVALC and create the numeric version as AVAL (ordered from best to worst response). This ordering is already covered within our RECIST v1.1 parameter derivation functions, and so changing AVAL here would not change the result of those derivations.

adrs <- adrs %>%
  mutate(
    AVALC = RSSTRESC,
    AVAL = aval_resp(AVALC)
  )

Flag Worst Assessment at Each Date (`ANL01FL`)

When deriving ANL01FL this is an opportunity to exclude any records that should not contribute to any downstream parameter derivations. In the below example this includes only selecting valid assessments and those occurring on or after randomization date. If there is more than one assessment at a date, the worst one is flagged.

worst_resp <- function(arg) {
  case_when(
    arg == "NE" ~ 1,
    arg == "CR" ~ 2,
    arg == "PR" ~ 3,
    arg == "SD" ~ 4,
    arg == "NON-CR/NON-PD" ~ 5,
    arg == "PD" ~ 6,
    TRUE ~ 0
  )
}

adrs <- adrs %>%
  restrict_derivation(
    derivation = derive_var_extreme_flag,
    args = params(
      by_vars = c(get_admiral_option("subject_keys"), exprs(ADT)),
      order = exprs(worst_resp(AVALC), RSSEQ),
      new_var = ANL01FL,
      mode = "last"
    ),
    filter = !is.na(AVAL) & ADT >= RANDDT
  )

Here is an alternative example where those records occurring after new anti-cancer therapy are additionally excluded (where NACTDT would be pre-derived as first date of new anti-cancer therapy. See admiralonco Creating and Using New Anti-Cancer Start Date for deriving this variable).

adrs <- adrs %>%
  mutate(
    ANL01FL = case_when(
      !is.na(AVAL) & ADT >= RANDDT & ADT < NACTDT ~ "Y",
      TRUE ~ NA_character_
    )
  )

Note here that we don’t filter out records after first PD at this stage, as that is specifically catered for in the admiralonco parameter derivation functions in the below steps, via source_pd arguments.

Flag Assessments up to First PD (`ANL02FL`)

However, if you prefer not to rely on source_pd arguments, then the user is free to filter out records after first PD at this stage in a similar way via a ANLzzFL flag, and then you could leave source_pd as null in all downstream parameter derivation function calls. So, for example the user could create ANL02FL flag to subset the post-baseline response data up to and including first reported progressive disease. This would be an alternative and transparent method to the use of source_pd argument approach to create ADRS parameters below. Using admiral function admiral::derive_var_relative_flag() we could create ANL02FL as below.

adrs <- adrs %>%
  derive_var_relative_flag(
    by_vars = get_admiral_option("subject_keys"),
    order = exprs(ADT, RSSEQ),
    new_var = ANL02FL,
    condition = AVALC == "PD",
    mode = "first",
    selection = "before",
    inclusive = TRUE
  )

Derive Progressive Disease Parameter

Now that we have the input records prepared above with any company-specific requirements, we can start to derive new parameter records. For the parameter derivations, all values except those overwritten by set_values_to argument are kept from the earliest occurring input record fulfilling the required criteria.

The function admiral::derive_extreme_records() can be used to find the date of first PD.

adrs <- adrs %>%
  derive_extreme_records(
    dataset_ref = adsl,
    dataset_add = adrs,
    by_vars = get_admiral_option("subject_keys"),
    filter_add = PARAMCD == "OVR" & AVALC == "PD" & ANL01FL == "Y",
    order = exprs(ADT, RSSEQ),
    mode = "first",
    exist_flag = AVALC,
    false_value = "N",
    set_values_to = exprs(
      PARAMCD = "PD",
      PARAM = "Disease Progression by Investigator",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

For progressive disease, response and death parameters shown in steps here and below, in our examples we show these as ADRS parameters, but they could equally be achieved via ADSL dates or ADEVENT parameters. If you prefer to store as an ADSL date, then the function admiral::derive_var_extreme_dt() could be used to find the date of first PD as a variable, rather than as a new parameter record. All the parameter derivation functions that use these dates are flexible to allow sourcing these from any input source using admiral::date_source(). See examples below.

Derive Response Parameter

The next required step is to define the source location for this newly derived PD date.

pd <- date_source(
  dataset_name = "adrs",
  date = ADT,
  filter = PARAMCD == "PD" & AVALC == "Y"
)

An equivalent example if using ADSL instead could be as follows (where PDDT would be pre-derived as first date of progressive disease).

pd <- date_source(
  dataset_name = "adsl",
  date = PDDT
)

The function derive_param_response() can then be used to find the date of first response. This differs from the admiral::derive_extreme_records() function in that it only looks for events occurring prior to first PD. In the below example, the response condition has been defined as CR or PR.

adrs <- adrs %>%
  derive_param_response(
    dataset_adsl = adsl,
    filter_source = PARAMCD == "OVR" & AVALC %in% c("CR", "PR") & ANL01FL == "Y",
    source_pd = pd,
    source_datasets = list(adrs = adrs),
    set_values_to = exprs(
      PARAMCD = "RSP",
      PARAM = "Response by Investigator (confirmation not required)",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

Derive Clinical Benefit Parameter

Similarly, we now define the source location for this newly derived first response date.

resp <- date_source(
  dataset_name = "adrs",
  date = ADT,
  filter = PARAMCD == "RSP" & AVALC == "Y"
)

The function derive_param_clinbenefit() can then be used to derive the clinical benefit parameter, which we define as a patient having had a response or a sustained period of time before first PD. This could also be known as disease control. In this example the “sustained period” has been defined as 42 days after randomization date, using the ref_start_window argument.

adrs <- adrs %>%
  derive_param_clinbenefit(
    dataset_adsl = adsl,
    filter_source = PARAMCD == "OVR" & ANL01FL == "Y",
    source_resp = resp,
    source_pd = pd,
    source_datasets = list(adrs = adrs),
    reference_date = RANDDT,
    ref_start_window = 42,
    set_values_to = exprs(
      PARAMCD = "CB",
      PARAM = "Clinical Benefit by Investigator (confirmation for response not required)",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

Derive Best Overall Response Parameter

The function derive_param_bor() can be used to derive the best overall response (without confirmation required) parameter. Similar to the above function you can optionally decide what period would you consider a SD or NON-CR/NON-PD as being eligible from. In this example, 42 days after randomization date has been used again.

adrs <- adrs %>%
  derive_param_bor(
    dataset_adsl = adsl,
    filter_source = PARAMCD == "OVR" & ANL01FL == "Y",
    source_pd = pd,
    source_datasets = list(adrs = adrs),
    reference_date = RANDDT,
    ref_start_window = 42,
    set_values_to = exprs(
      PARAMCD = "BOR",
      PARAM = "Best Overall Response by Investigator (confirmation not required)",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = aval_resp(AVALC),
      ANL01FL = "Y"
    )
  )

Note that the above gives pre-defined AVAL values (defined by aval_resp()) of: "CR" ~ 1, "PR" ~ 2, "SD" ~ 3, "NON-CR/NON-PD" ~ 4, "PD" ~ 5, "NE" ~ 6, "MISSING" ~ 7.

If you’d like to provide your own company-specific ordering here you could do this as follows:

aval_resp_new <- function(arg) {
  case_when(
    arg == "CR" ~ 7,
    arg == "PR" ~ 6,
    arg == "SD" ~ 5,
    arg == "NON-CR/NON-PD" ~ 4,
    arg == "PD" ~ 3,
    arg == "NE" ~ 2,
    arg == "MISSING" ~ 1,
    TRUE ~ NA_real_
  )
}

Then update the definition of AVAL in the set_values_to argument of the above derive_param_bor() call. Be aware that this will only impact the AVAL mapping, not the derivation of BOR in any way - as the function derivation relies only on AVALC here.

Derive Best Overall Response of CR/PR Parameter

The function admiral::derive_extreme_records() can be used to check if a patient had a response for BOR.

adrs <- adrs %>%
  derive_extreme_records(
    dataset_ref = adsl,
    dataset_add = adrs,
    by_vars = get_admiral_option("subject_keys"),
    filter_add = PARAMCD == "BOR" & AVALC %in% c("CR", "PR"),
    order = exprs(ADT, RSSEQ),
    mode = "first",
    exist_flag = AVALC,
    false_value = "N",
    set_values_to = exprs(
      PARAMCD = "BCP",
      PARAM = "Best Overall Response of CR/PR by Investigator (confirmation not required)",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

Derive Response Parameters requiring Confirmation

Any of the above response parameters can be repeated for “confirmed” responses only. For these the functions derive_param_confirmed_resp() and derive_param_confirmed_bor() can be used. Some of the other functions from above can then be re-used passing in these confirmed response records. See the examples below of derived parameters requiring confirmation. The assessment and the confirmatory assessment here need to occur at least 28 days apart (without any +1 applied to this calculation of days between visits), using the ref_confirm argument.

adrs <- adrs %>%
  derive_param_confirmed_resp(
    dataset_adsl = adsl,
    filter_source = PARAMCD == "OVR" & ANL01FL == "Y",
    source_pd = pd,
    source_datasets = list(adrs = adrs),
    ref_confirm = 28,
    set_values_to = exprs(
      PARAMCD = "CRSP",
      PARAM = "Confirmed Response by Investigator",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

confirmed_resp <- date_source(
  dataset_name = "adrs",
  date = ADT,
  filter = PARAMCD == "CRSP" & AVALC == "Y"
)

adrs <- adrs %>%
  derive_param_clinbenefit(
    dataset_adsl = adsl,
    filter_source = PARAMCD == "OVR" & ANL01FL == "Y",
    source_resp = confirmed_resp,
    source_pd = pd,
    source_datasets = list(adrs = adrs),
    reference_date = RANDDT,
    ref_start_window = 42,
    set_values_to = exprs(
      PARAMCD = "CCB",
      PARAM = "Confirmed Clinical Benefit by Investigator",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  ) %>%
  derive_param_confirmed_bor(
    dataset_adsl = adsl,
    filter_source = PARAMCD == "OVR" & ANL01FL == "Y",
    source_pd = pd,
    source_datasets = list(adrs = adrs),
    reference_date = RANDDT,
    ref_start_window = 42,
    ref_confirm = 28,
    set_values_to = exprs(
      PARAMCD = "CBOR",
      PARAM = "Best Confirmed Overall Response by Investigator",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = aval_resp(AVALC),
      ANL01FL = "Y"
    )
  ) %>%
  derive_extreme_records(
    dataset_ref = adsl,
    dataset_add = adrs,
    by_vars = get_admiral_option("subject_keys"),
    filter_add = PARAMCD == "CBOR" & AVALC %in% c("CR", "PR"),
    order = exprs(ADT, RSSEQ),
    mode = "first",
    exist_flag = AVALC,
    false_value = "N",
    set_values_to = exprs(
      PARAMCD = "CBCP",
      PARAM = "Best Confirmed Overall Response of CR/PR by Investigator",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

Derive Parameters using Independent Review Facility (IRF)/Blinded Independent Central Review (BICR) responses

All of the above steps can be repeated for different sets of records, such as now using assessments from the IRF/BICR instead of investigator. For this you would just need to replace the first steps with selecting the required records, and then feed these as input to the downstream parameter functions.

Remember that a new progressive disease and response source object would be required for passing to source_pd and source_resp respectively.

adrs_bicr <- rs %>%
  filter(
    RSEVAL == "INDEPENDENT ASSESSOR" & RSACPTFL == "Y" & RSTESTCD == "OVRLRESP"
  ) %>%
  mutate(
    PARAMCD = "OVRB",
    PARAM = "Overall Response by BICR",
    PARCAT1 = "Tumor Response",
    PARCAT2 = "Blinded Independent Central Review",
    PARCAT3 = "RECIST 1.1"
  )

Then in all the calls to the parameter derivation functions you would replace the PARAMCD == "OVR" source with PARAMCD == "OVRR1".

Derive Death Parameter

The function admiral::derive_extreme_records() can be used to create a new death parameter using death date from ADSL. We need to restrict the columns from ADSL as we’ll merge all required variables later across all our ADRS records.

adsldth <- adsl %>%
  select(!!!get_admiral_option("subject_keys"), DTHDT, !!!adsl_vars)

adrs <- adrs %>%
  derive_extreme_records(
    dataset_ref = adsldth,
    dataset_add = adsldth,
    by_vars = get_admiral_option("subject_keys"),
    filter_add = !is.na(DTHDT),
    exist_flag = AVALC,
    false_value = "N",
    set_values_to = exprs(
      PARAMCD = "DEATH",
      PARAM = "Death",
      PARCAT1 = "Reference Event",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y",
      ADT = DTHDT
    )
  ) %>%
  select(-DTHDT)

Derive Last Disease Assessment Parameters

The function admiral::derive_extreme_records() can be used to create a parameter for last disease assessment.

adrs <- adrs %>%
  derive_extreme_records(
    dataset_ref = adsl,
    dataset_add = adrs,
    by_vars = get_admiral_option("subject_keys"),
    filter_add = PARAMCD == "OVR" & ANL01FL == "Y",
    order = exprs(ADT, RSSEQ),
    mode = "last",
    set_values_to = exprs(
      PARAMCD = "LSTA",
      PARAM = "Last Disease Assessment by Investigator",
      PARCAT1 = "Tumor Response",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      ANL01FL = "Y"
    )
  )

Derive Measurable Disease at Baseline Parameter

The function admiral::derive_param_exist_flag() can be used to check whether a patient has measurable disease at baseline, according to a company-specific condition. In this example we check TU for target lesions during the baseline visit. We need to restrict the columns from ADSL as we’ll merge all required variables later across all our ADRS records.

adslmdis <- adsl %>%
  select(!!!get_admiral_option("subject_keys"), !!!adsl_vars)

adrs <- adrs %>%
  derive_param_exist_flag(
    dataset_ref = adslmdis,
    dataset_add = tu,
    condition = TUEVAL == "INVESTIGATOR" & TUSTRESC == "TARGET" & VISIT == "SCREENING",
    false_value = "N",
    missing_value = "N",
    set_values_to = exprs(
      PARAMCD = "MDIS",
      PARAM = "Measurable Disease at Baseline by Investigator",
      PARCAT2 = "Investigator",
      PARCAT3 = "RECIST 1.1",
      AVAL = yn_to_numeric(AVALC),
      ANL01FL = "Y"
    )
  )

Assign `ASEQ`

The function admiral::derive_var_obs_number() can be used to derive ASEQ. An example call is:

adrs <- adrs %>%
  derive_var_obs_number(
    by_vars = get_admiral_option("subject_keys"),
    order = exprs(PARAMCD, ADT, VISITNUM, RSSEQ),
    check_type = "error"
  )

Add ADSL variables

If needed, the other ADSL variables can now be added. List of ADSL variables already merged held in vector adsl_vars.

adrs <- adrs %>%
  derive_vars_merged(
    dataset_add = select(adsl, !!!negate_vars(adsl_vars)),
    by_vars = get_admiral_option("subject_keys")
  )

Example Script

ADaM	Sample Code
ADRS_BASIC	`admiral::use_ad_template("ADRS_BASIC", package = "admiralonco")`