admiral-adrs

name	admiral-adrs
description	Derives an ADaM Tumor Response Analysis Dataset (ADRS) using the {admiral} and {admiralonco} R packages. Use when a user needs to create ADRS from SDTM RS domain data, derive RECIST 1.1 response parameters (overall response, confirmed response, best overall response, clinical benefit), or generate QC-ready R code following CDISC ADaM and oncology conventions. Requires SDTM RS input data, ADSL with randomization and treatment dates, and an ADaM ADRS specification.
license	MIT
metadata	{"author":"Navitas Data Sciences","version":"0.1","pharmaverse":"true","parent":"admiral"}
compatibility	Requires R with admiral, admiralonco, dplyr, lubridate, and pharmaversesdtm installed. Requires a completed ADSL dataset with RANDDT and TRTSDT. Designed for use in a GxP-compliant oncology trial environment with access to SDTM RS domain data and an ADaM ADRS specification.

admiral-adrs

Shared conventions (library setup, pipe style, date rules, flag convention, # REVIEW: annotations, stopifnot() patterns) are defined in the parent ../SKILL.md. The workflow below is ADRS-specific.

Derives a CDISC-conformant ADRS tumor response dataset using {admiral} and {admiralonco}. Outputs executable, QC-ready R code covering RECIST 1.1 response parameters with full derivation traceability.

The primary design challenge in ADRS is the confirmation logic: CR and PR must be supported by a second qualifying assessment ≥28 days later. Best Overall Response (BOR) follows a strict hierarchy (CR > PR > SD > NON-CR/NON-PD > PD > NE) and handles NE propagation in ways that manual case_when() or slice_min() cannot replicate correctly. Always use admiralonco functions — never manual derivations.

---

Inputs

Before generating code, confirm the following are available or explicitly noted as absent:

Input	Required	Notes
RS	Yes	One record per tumor assessment per subject; RSSTRESC contains the response category (CR/PR/SD/PD/NE)
ADSL	Yes	Provides TRTSDT, RANDDT, treatment labels, population flags
ADaM ADRS spec	Yes	Parameter list, PARAMCD/PARAM mapping, confirmation window, clinical benefit definition
Study context	Yes	RECIST version (1.0 vs 1.1), assessor (investigator vs BICR), confirmation window, clinical benefit anchor date

If RS or ADSL are absent, stop and request them.

Note on pharmaversesdtm test data: The pharmaversesdtm package no longer exports a plain rs object. Use pharmaversesdtm::rs_onco_recist for test or benchmark runs. When working with real study data, load RS from the study's SDTM package or file path.

---

Workflow

Follow these steps in order. Generate code section by section, not as a single block.

Step 1 — Setup and domain loading

library(admiral)
library(admiralonco)
library(dplyr)
library(lubridate)
library(pharmaversesdtm)

# Load RS domain
# For pharmaversesdtm test data: use rs_onco_recist (plain `rs` no longer exported)
rs   <- pharmaversesdtm::rs_onco_recist
adsl <- adsl  # assumed derived upstream; replace with path/load as needed

# Confirm RS has at least one record
stopifnot(nrow(rs) > 0)

Step 2 — DOMAIN removal

Remove DOMAIN from RS before any derive_param_*() or derive_vars_merged() calls. admiral errors when DOMAIN exists in both the dataset and a source dataset passed to these functions.

rs <- rs |> select(-DOMAIN)

Step 3 — Merge ADSL backbone variables

Bring required ADSL variables into the RS dataset before parameter derivation. At minimum: RANDDT and TRTSDT (clinical benefit anchor and study day reference), treatment labels, and population flags.

# REVIEW: Confirm which ADSL variables are required by the ADaM ADRS spec.
#   RANDDT is the conventional clinical benefit anchor date; TRTSDT is required
#   for ADY derivation. Add or remove population flags per the spec.
adrs <- rs |>
  derive_vars_merged(
    dataset_add = adsl,
    by_vars     = exprs(STUDYID, USUBJID),
    new_vars    = exprs(RANDDT, TRTSDT, TRT01P, TRT01PN, TRT01A, TRT01AN,
                        ITTFL, SAFFL)
  )

Step 4 — Date derivation (ADT, ADTF, ADY)

Derive the analysis date from RSDTC. This must happen before any derive_param_*() call — admiralonco response functions use ADT internally for confirmation window comparisons.

adrs <- adrs |>
  derive_vars_dt(
    dtc             = RSDTC,
    new_vars_prefix = "A",
    date_imputation = "first",
    flag_imputation = "auto"
  ) |>
  derive_vars_dy(
    reference_date = TRTSDT,
    source_vars    = exprs(ADT)
  )

Step 5 — Analysis value (AVALC, AVAL)

Set AVALC from RSSTRESC and derive the numeric AVAL using the admiralonco helper aval_resp(). This function maps response categories to a monotone numeric scale: CR=1, PR=2, SD=3, NON-CR/NON-PD=4, PD=5, NE=6.

# REVIEW: Confirm that RSSTRESC values in this study's RS domain align with the
#   RECIST 1.1 CT expected by aval_resp(). Non-standard categories (e.g.
#   "NON-CR/NON-PD" in lymphoma, iRECIST response categories) require a custom
#   lookup if aval_resp() does not map them — do not suppress the resulting NA.
adrs <- adrs |>
  mutate(
    AVALC = RSSTRESC,
    AVAL  = aval_resp(AVALC)
  )

Step 6 — Overall response parameter (OVRLRESP)

Add OVRLRESP records — one per subject per assessment timepoint. This parameter carries the verbatim response at each visit; it is not confirmation-adjusted.

# REVIEW: PARAMCD and PARAM values must match the ADaM ADRS spec exactly.
#   Adjust filter_source if the study uses a different assessor
#   (BICR: RSEVAL == "INDEPENDENT ASSESSOR") or a different RECIST version.
adrs <- adrs |>
  derive_param_response(
    dataset_adsl  = adsl,
    filter_source = RSEVAL == "INVESTIGATOR" & RSEVALID == "RECIST 1.1",
    source_var    = RSSTRESC,
    set_values_to = exprs(
      PARAMCD = "OVRLRESP",
      PARAM   = "Overall Response by Investigator"
    )
  )

Step 7 — Unconfirmed response flag (RSP)

Add RSP records — one per subject, indicating whether any CR or PR was observed (unconfirmed). Retained alongside CONFIRMED to enable the ORR verification check in Step 11.

adrs <- adrs |>
  derive_param_response(
    dataset_adsl  = adsl,
    filter_source = RSEVAL == "INVESTIGATOR" & RSEVALID == "RECIST 1.1",
    source_var    = RSSTRESC,
    resp_val      = c("CR", "PR"),
    set_values_to = exprs(
      PARAMCD = "RSP",
      PARAM   = "Response (Unconfirmed)"
    )
  )

Step 8 — Confirmed response (CONFIRMED)

A CR or PR is confirmed when a second qualifying assessment ≥ ref_confirm days after the first also shows CR or PR. CONFIRMED = "Y" if the subject has at least one confirmed CR or PR; "N" otherwise.

Flag convention note: CONFIRMED uses "Y"/"N", not "Y"/NA — this is the documented admiralonco contract. See Flag convention exception below. Do not recode "N" to NA.

# REVIEW: ref_confirm = 28 is the RECIST 1.1 standard for CR/PR confirmation.
#   Confirm this value against the study protocol and SAP — some programs specify
#   21 days, and regulatory precedent exists for other windows. A wrong value
#   silently inflates or deflates confirmed ORR.
adrs <- adrs |>
  derive_param_confirmed_resp(
    dataset_adsl  = adsl,
    filter_source = RSEVAL == "INVESTIGATOR" & RSEVALID == "RECIST 1.1",
    source_var    = RSSTRESC,
    ref_confirm   = 28,       # PLACEHOLDER — confirm from SAP
    set_values_to = exprs(
      PARAMCD = "CONFIRMED",
      PARAM   = "Confirmed Response"
    )
  )

Step 9 — Best overall response (BESTRESP)

BOR applies the RECIST 1.1 hierarchy across all assessments for each subject. derive_param_confirmed_bor() handles confirmation requirements, the NE propagation rule, and the hierarchy correctly. Never substitute slice_min(AVAL) or manual case_when() — these cannot replicate the NE propagation logic.

# REVIEW: missing_as_ne controls how missing assessments are treated in BOR.
#   FALSE (default): missing assessments are ignored (excluded from BOR).
#   TRUE: missing assessments count as NE, which can worsen BOR for subjects
#   with gaps in the assessment schedule.
#   Confirm the per-protocol analysis population definition with the
#   statistical reviewer before committing to either value.
adrs <- adrs |>
  derive_param_confirmed_bor(
    dataset_adsl  = adsl,
    filter_source = RSEVAL == "INVESTIGATOR" & RSEVALID == "RECIST 1.1",
    source_var    = RSSTRESC,
    ref_confirm   = 28,       # must match Step 8
    missing_as_ne = FALSE,    # PLACEHOLDER — confirm from SAP
    set_values_to = exprs(
      PARAMCD = "BESTRESP",
      PARAM   = "Best Overall Response (Confirmed)"
    )
  )

Step 10 — Clinical benefit (CBRESPFL)

Clinical benefit is a durable non-progressive response: CR, PR, or SD sustained for ≥ ref_start_window days from reference_date. CBRESPFL = "Y" if criteria are met; "N" otherwise.

Flag convention note: CBRESPFL uses "Y"/"N", not "Y"/NA. Same admiralonco contract as CONFIRMED — do not recode.

# REVIEW: Two protocol-specific decisions are required here and must both be
#   confirmed from the protocol and SAP before use:
#
#   (1) reference_date — RANDDT is the conventional anchor. Some protocols
#       instead define the 42-day window from the date of the first qualifying
#       response (first SD, PR, or CR). If the protocol means "from first
#       qualifying assessment", compute per-subject first-assessment dates
#       and pass them as reference_date rather than a fixed ADSL variable.
#
#   (2) ref_start_window = 42 — RECIST standard for SD duration. Some studies
#       use 35 or 56 days. Confirm from the protocol.
adrs <- adrs |>
  derive_param_clinbenefit(
    dataset_adsl     = adsl,
    filter_source    = RSEVAL == "INVESTIGATOR" & RSEVALID == "RECIST 1.1",
    source_var       = RSSTRESC,
    reference_date   = RANDDT,    # PLACEHOLDER — confirm anchor from protocol
    ref_start_window = 42,        # PLACEHOLDER — confirm from protocol
    set_values_to    = exprs(
      PARAMCD = "CBRESPFL",
      PARAM   = "Clinical Benefit"
    )
  )

Step 11 — Verification: confirmed vs unconfirmed ORR

Print response counts side-by-side and assert that confirmed ORR cannot exceed unconfirmed ORR. A confirmed count greater than unconfirmed count signals a configuration error in the confirmation window.

# Print response parameter summary for QC — include in script output log
response_summary <- adrs |>
  filter(PARAMCD %in% c("RSP", "CONFIRMED", "CBRESPFL")) |>
  count(PARAMCD, AVALC)
print(response_summary)

# Confirmed ORR must not exceed unconfirmed ORR
n_confirmed   <- sum(adrs$PARAMCD == "CONFIRMED" & adrs$AVALC == "Y", na.rm = TRUE)
n_unconfirmed <- sum(adrs$PARAMCD == "RSP"       & adrs$AVALC == "Y", na.rm = TRUE)
stopifnot(n_confirmed <= n_unconfirmed)

# PD is never a confirmed response
stopifnot(
  adrsPARAMCD    adrsAVALC   
  adrsAVAL  aval_resp
  na.rm  

Step 12 — Final checks

# Required parameter coverage
required_params <- c("OVRLRESP", "RSP", "CONFIRMED", "BESTRESP", "CBRESPFL")
missing_params  <- setdiff(required_params, unique(adrs$PARAMCD))
if (length(missing_params) > 0) {
  stop("Missing required ADRS parameters: ", paste(missing_params, collapse = ", "))
}

# Uniqueness: one record per subject per subject-level parameter
dup_check <- adrs |>
  filter(PARAMCD %in% c("BESTRESP", "CONFIRMED", "RSP", "CBRESPFL") 
  countSTUDYID USUBJID PARAMCD 
  filtern  
 nrowdup_check   
  stop
       pastepastedup_checkUSUBJID dup_checkPARAMCD collapse  



required_vars  
     
       

missing_vars  setdiffrequired_vars adrs
 missing_vars   
  stop pastemissing_vars collapse  

---

Flag convention exception for CONFIRMED and CBRESPFL

The admiral family convention is "Y" or NA — never "N" — for flag variables. ADRS has two named exceptions:

Variable	Values	Reason
CONFIRMED	"Y" / "N"	admiralonco contract: "N" means assessed but unconfirmed, not missing
CBRESPFL	"Y" / "N"	admiralonco contract: same logic as CONFIRMED

Do not recode "N" to NA for these variables. The "N" records are required by downstream derive_param_confirmed_bor() logic to correctly identify subjects with no confirmed response. Recoding them breaks BOR derivation.

All other flags in ADRS (ANL01FL, ABLFL, population flags from ADSL) follow the standard "Y" or NA convention.

---

Common errors to avoid

• Using rs instead of rs_onco_recist from pharmaversesdtm — the plain rs object is no longer exported; the script will fail at load
• Using case_when() or slice_min(AVAL) for BOR — manual BOR derivation cannot replicate admiralonco's NE propagation rule and confirmation hierarchy; always use derive_param_confirmed_bor()
• Not removing DOMAIN from RS before derive_param_*() calls — causes variable conflict errors; DOMAIN must be removed in Step 2, not in the final select() at the end of the script
• Hardcoding ref_confirm = 28 without a # REVIEW: comment — the confirmation window is protocol-specific and must come from the SAP; a wrong value silently inflates or deflates confirmed ORR
• Applying derive_vars_dt() after derive_param_response() — ADT must exist before the response parameter functions run; the functions use ADT for confirmation window comparisons
• Using RANDDT as the clinical benefit anchor without verifying protocol intent — if the protocol defines the 42-day window from the "first qualifying response assessment" rather than randomization, a fixed RANDDT will misclassify subjects
• Not printing the confirmed vs unconfirmed ORR comparison — the omission is a silent audit risk; confirmed ORR > unconfirmed ORR indicates a derivation error that will not surface without an explicit check
• Recoding CONFIRMED or CBRESPFL "N" to NA — these values are meaningful per the admiralonco function contract; recoding breaks downstream BOR derivation

---

Output checklist

Before returning code, verify:

• DOMAIN removed from RS in Step 2, before any derive_param_*() call
• ADSL merged with at minimum RANDDT, TRTSDT, and population flags before parameter derivation
• ADT derived with derive_vars_dt() from RSDTC, with flag_imputation = "auto"
• AVALC and AVAL (via aval_resp()) populated before derive_param_response() calls
• All five parameters present in output: OVRLRESP, RSP, CONFIRMED, BESTRESP, CBRESPFL
• # REVIEW: at PARAMCD/PARAM mapping in Steps 6–10
• # REVIEW: at ref_confirm in Steps 8 and 9
• # REVIEW: at missing_as_ne in Step 9
• # REVIEW: at reference_date and ref_start_window in Step 10
• Confirmed vs unconfirmed ORR comparison printed to console (Step 11)
• stopifnot(n_confirmed <= n_unconfirmed) present (Step 11)
• PD-never-confirmed assertion present (Step 11)
• Required PARAMCD coverage check present (Step 12)
• Uniqueness assertion for subject-level parameters (Step 12)
• CONFIRMED and CBRESPFL values are "Y"/"N" — not recoded to "Y"/NA