trialdiff

trialdiff is a clinical-trial data-cut change detection and downstream impact assessment framework for R, designed for the pharmaverse ecosystem.

Existing tools tell you that two datasets differ. trialdiff answers the clinical-programming question that follows:

What changed, what does the change represent, and which downstream analyses or outputs might be affected?

It does this in five transparent, composable layers:

Layer	Function	Purpose
Compare	`compare_cut()`	Added/removed/modified observations and schema changes
Classify	`classify_changes()`	Rule-based clinical change taxonomy
Trace	`define_lineage()`, `trace_dependencies()`, `lineage_from_metadata()`, `output_registry()`	Explicit data lineage graph
Assess	`assess_impact()`	Definitely / potentially / unlikely impact
Report	`report_diff()`	HTML, Quarto and machine-readable JSON

Everything is deterministic. There is no machine learning, and no statistical impact is ever claimed - analyses are flagged for review and rerun.

Installation

# From r-universe (includes Windows/macOS binaries)
install.packages(
  "trialdiff",
  repos = c(
    hirujan = "https://hirujan-r.r-universe.dev",
    CRAN = "https://cloud.r-project.org"
  )
)

# Or from GitHub
# install.packages("remotes")
remotes::install_github("Hirujan-R/trialdiff")

Quick start

library(trialdiff)

diff <- compare_cut(
  old = adsl_cut1,
  new = adsl_cut2,
  by = "USUBJID",
  dataset = "ADSL"
)

classified <- classify_changes(diff)

impact <- assess_impact(classified, adsl_adlb_lineage)

report_diff(diff, impact = impact, output = "report.html")

A treatment assignment change (TRT01P = "Placebo" -> "Drug A") is detected, classified as a treatment-assignment change, traced through ADSL.TRT01P -> ADLB.TRT01P -> lab summary -> MMRM -> efficacy table, and every downstream object is flagged for review with a rationale.

Why not just `diffdf`?

diffdf (and waldo) are excellent low-level comparison tools, and trialdiff deliberately does not reinvent them. trialdiff adds the layers that are specific to clinical programming:

a clinical change taxonomy (new subject, new visit, corrected value, missingness transitions, derived-variable change, treatment-assignment change, metadata change, …);
explicit lineage between datasets, variables, analyses and outputs;
auditable impact assessment with a documented policy;
review reports that name the objects a programmer or statistician must check.

It can even delegate the low-level comparison itself: compare_cut(..., backend = "diffdf") uses diffdf to detect differences and translates the result into the same tdiff object.

Design principles

Deterministic comparison, not opaque inference.
Transparent, user-extensible rules.
Explicit, reproducible lineage metadata.
Auditable impact assessment with a written rationale for every flag.
Clinically meaningful reporting.

Documentation

vignette("trialdiff") - getting started.
vignette("change-classification") - the rule system.
vignette("lineage-and-impact") - lineage and impact.
vignette("ecosystem") - relationship to diffdf, admiral, metacore, cards, tern, SAS PROC COMPARE and others.
vignette("case-study") - end-to-end walkthrough on public pharmaverseadam data.

Project proposal

A full proposal covering the problem statement, ecosystem/gap analysis, novelty assessment, architecture, testing and roadmap is available in the project proposal.

Code of conduct

Please note that the trialdiff project is released with a Contributor Code of Conduct. By contributing you agree to abide by its terms.

License

MIT (c) Hirujan Rangaraj. No proprietary or real patient data is included; all example datasets are synthetic.