[benchmark][r2rtf] Kaplan-Meier survival analysis table — median survival, confidence intervals, number at risk, and censoring annotation

[efchea1]

This benchmark extends the r2rtf skill beyond baseline characteristics (Issue #128) to survival analysis output, the most common efficacy deliverable in oncology and cardiovascular submissions. A KM table with median survival, confidence intervals, number at risk, and log-rank p-value is required in virtually every oncology NDA and BLA. The silent failure tested here produces a table with correct numbers but missing or incorrectly formatted required elements that FDA statistical reviewers' flag on initial document review.

Skills

r2rtf

Language (Optional)

R

Query

Using the pharmaverseadam ADTTE dataset, produce a submission-ready RTF survival analysis table for a DMC or regulatory submission.

Input data:
library(pharmaverseadam)
adtte <- pharmaverseadam::adtte

The table must include:

1. Kaplan-Meier estimates by treatment arm (TRT01P)
2. Median survival with 95% CI, use Brookmeyer-Crowley method if available (km.ci package); if not available, use survfit() with conf.type = "log"and document the method used with a

   REVIEW: flag for biostatistician confirmation

3. Number at risk at: 0, 3, 6, 12, 18, 24 months
4. Log-rank test p-value (unstratified, from survdiff())
5. Hazard ratio with 95% CI from coxph()
6. Censoring footnote if any CNSR == 1 records present

Formatting requirements:

• Median survival: "X.X (Y.Y, Z.Z)", months, one decimal place
• CI notation: parentheses, brackets are a failing answer
• Number at risk: integers only, decimal values indicate wrong risk set computation
• HR: "X.XX (Y.YY, Z.ZZ)", two decimal places
• P-value: "< 0.001" when p < 0.001, "0.000" is a failing answer
• Title, footnote defining CI/HR/NR, page number in footer
• RTF file written to disk using write_rtf()

Expected Output

Executable R code using {r2rtf} pipe API producing a valid .rtf file. The script must call write_rtf() to produce a file on disk, displaying a data frame in the console is not acceptable.

Critical note: Five silent failures are common in KM table production: (1) Missing number-at-risk row, FDA requires this to assess tail estimate reliability; (2) Wrong CI method without documentation, Brookmeyer-Crowley is preferred; if another method is used it must be explicitly flagged for review; (3) Wrong CI notation, parentheses required, not brackets; (4) P-value "0.000", must display as "< 0.001"; (5) Fractional at-risk counts, number at risk must be non-negative integers. All five pass visual inspection but fail formal QC review.

Attached Files / Input Context (Optional)

library(pharmaverseadam)
adtte <- pharmaverseadam::adtte

Rubric Criteria (Assertions)

r2rtf API correctness

• rtf_body(), rtf_title(), rtf_footnote(), rtf_page() used in pipe chain, manual RTF string construction absent
• write_rtf() called and produces .rtf file on disk, console output only is a failing answer
• RTF file opens in Microsoft Word without corruption warnings

Statistical correctness

• Median survival CI uses Brookmeyer-Crowley (km.ci::km.ci()) OR survfit() with explicit conf.type documented and # REVIEW: flag, default log CI without documentation is a failing answer
• Log-rank p-value from survdiff() — t-test or Wilcoxon absent
• HR and CI from coxph() — manual computation absent
• Number at risk from summary(survfit_object, times = c(0,3,6,12,18,24))

Formatting correctness

• Median CI in parentheses: "X.X (Y.Y, Z.Z)", brackets absent
• P-value "< 0.001" when p < 0.001, "0.000" or "0.0000" absent
• HR to two decimals: "X.XX (Y.YY, Z.ZZ)"
• Number at risk values are non-negative integers, decimal values indicate incorrect risk set computation and are a failing answer
• Number at risk row present and aligned with month columns
• Censoring footnote present if any CNSR == 1 records in ADTTE

Submission compliance

• Title includes analysis description (not blank or placeholder)
• Footnote defines: CI = confidence interval, HR = hazard ratio, NR = not reached (or equivalent)
• Page number in document footer
• Treatment arm labels from TRT01P, not hardcoded strings

QC readiness

• REVIEW: at CI method (confirm Brookmeyer-Crowley vs log-transformed, document in SAP before submission)

• REVIEW: at p-value threshold (confirm "< 0.001" display against SAP formatting conventions)

• REVIEW: at number-at-risk timepoints (confirm 0, 3, 6, 12,18, 24 months against DMC charter or SAP)

• REVIEW: at Cox model (confirm unstratified vs stratified against SAP randomization strata)

• RTF file produced on disk and opens without errors
• Script runs end-to-end without errors