Trial biostatistics toolkit

Open-source tools, templates and worked examples for biostatistics consultants and employees across study planning, programming, operations and reporting. The worked examples focus on early-phase clinical pharmacology; each tool can be adapted where its methods and input contract fit the study.

Choose your task and get started · Home · Full library

Point-and-click setup for the whole toolkit · Control panel and module catalog · Download the complete ZIP

For local workflows, download and extract the ZIP, then double-click Start Toolkit.command (macOS) or Start Toolkit.cmd (Windows). The local browser control panel covers all five Shiny apps, the reproducible R examples, QC checks, browser workspaces, paired SAS/R libraries and configuration templates. Create a private project from examples for each study or engagement, check dependencies, install supported R packages, open apps, upload documented inputs and export the project through the GUI. Python 3.10+ starts the control panel; each card identifies any additional runtime installer. The Linux launcher is Start Toolkit.sh.

The toolkit_launcher/ catalog distinguishes runnable examples from templates and references. The four published Shinylive apps and browser-only workspaces run without local installation. The vaccine app accepts mapped study CSVs and computes results entirely in the browser. The external eCRF apps remain a separate repository. The shared tracker setup handles its accounts, hosting and optional email.

Run vaccine analyses in your browser: no R installation, account or toolkit download. Start with a blank workspace or deliberately load the synthetic example, map CSVs, confirm study definitions, validate and run all six same-core R outputs. Download the complete results ZIP, individual tables/figures, setup profile and configured SAS reference. Settings or input changes invalidate earlier results; reset or reload clears the browser session. Includes an illustrated learning companion. Browser limits and workflow are documented separately from the optional desktop route; native SAS remains unexecuted.

Guided template setup: choose study design and outputs, upload and map CSV inputs, preview requirements, then generate a versioned study package. Forms cover the SAS/R TLF libraries, monitoring checks, LOCALMIND findings and AE/PK draft reports. Save or import setup profiles to reuse choices across projects. The 221 R TLF programs, eight monitoring checks and aggregate interpretation verification can then run through buttons. Each attempt retains its own complete results and QC report. R is the runnable default; SAS reference exports remain available.

This is a general-purpose toolkit, not a deployment for a specific client. A new private project contains fictional starter data, not an empty or already configured production study. Source files, generated results and logs are retained in that project. Export app/session data before closing; a project ZIP does not capture unsaved Shiny state or browser storage. The shared tracker’s database and settings live outside that workspace ZIP; use its separate backup controls. Native SAS reference code has not been executed. Study-specific scientific review and model-service configuration remain separate. See the setup guide for inputs, persistence, runtime installation and all 23 module paths, and QC_REPORT.md for the October 2026 QC scope.

Rscript launch_app.R --list
Rscript launch_app.R --check
Rscript launch_app.R trialmon --install
# Also: lifecycle, uat, risk, hours.

The R launcher finds the complete checkout from another working directory. Package downloads happen only with --install and target a user library.

Site: https://patrickiben.github.io/biometrics-ai-portfolio/

Author: Patrick Iben

---

What's inside

Statistical programming (SAS + R)

• ADaM-based TLF libraries: 221 Table/Listing/Figure programs each in SAS and R (pharmaverse: dplyr + rtables + ggplot2 + emmeans), across every early-phase design (parallel, crossover, single/fixed-sequence, SAD, MAD). Geometric PK on the log scale, distinct-participant counts, period denominators. The SAS programs have not been executed (no SAS license); the R programs parse but have not been run on data. → sas_tlf_pseudocode/ · r_tlf_pseudocode/
• Worked bioequivalence study: one 2x2 crossover taken from PowerTOST sample size and a blockrand schedule through an ICH E9(R1) estimand and SAP excerpt, admiral ADaM with a Pinnacle 21-layout specification and XPT files, PKNCA, the ICH M13A analysis, rtables/tern tables with RTF, a QC re-derivation by a second method and CSR section 11-style text. One script (study/run_all.R) regenerates every output and CI reruns it. QC is a same-author re-derivation, not independent double programming. → study/
• Synthetic ADaM generator + study-lifecycle sandbox: a self-contained R generator (one-compartment oral PK + NCA, a 2x2 bioequivalence crossover, per-period safety domains) that emits 8 ADaM domains and runs demographics, PK/NCA and the bioequivalence ANOVA end to end. → sim_lifecycle/
• Operating-characteristics studies: residual carryover and the ICH M13A 5% pre-dose rule (sim_lifecycle/carryover_oc/), and a prespecified C-QTc hysteresis check with its size and power (cqtc_hysteresis/).
• Protocol→CSR deliverables pipeline and a (a small-model skeleton that never sources a number; the model step is a stub and the number validator runs). → ·

Operational tooling (no AI required)

• Deterministic SAS/R monitoring + coverage for leave and handoffs: configuration templates for scheduled checks, evidence packets and notifications. Paths, schedules and recipients require local configuration; the templates do not detect sensitive content or enforce an organization's data-sharing policy. → sas_r_automation/ · sasr_monitoring_wiki/
• Coverage and continuity toolkit: the leave playbook generalized to five scenarios (planned leave, unplanned absence, role handoff or departure, a new owner taking over, bus-factor review), plus an interactive single-file page: enter or import the recurring jobs you run for a client (owner, backups, schedule, inputs, outputs, criticality, last run, runbook) and get a coverage matrix, uncovered and single-owner risks, a checklist per job, and a printable or downloadable handoff packet. Data stays in the browser (localStorage + JSON/CSV export); no network requests; ships with a clearly fictional example. → sasr_onleave_wiki/ · playbook · CSV template
• Open Study Tracker: an MIT-licensed deliverables tracker with local SQLite storage, keyed CSV/JSON imports, dry runs and change history. Use the offline browser workspace to review and edit status, then exchange files with scheduled R/SAS jobs. The offline version needs no account or API key. Optional self-hosted shared mode adds individual logins, viewer/editor permissions, conflicting-edit protection, and queued SMTP alerts with retries and daily exception digests. → open_tracker/ · file workflow ·

Interactive apps and dashboards

• Participant-safety monitoring (TRIALMON, with a Safety Review Committee packet), trial-risk early warning (participant and study tiers), and a study-lifecycle monitor. These include synthetic examples for exploration. → sasr_monitoring_wiki/TRIALMON_Dashboard.html · risk_monitor/ · lifecycle_dashboard/
• The TRIALMON, UAT and lifecycle R Shiny apps also run in the browser with Shinylive: apps/ (TRIALMON) · UAT tracker · lifecycle monitor.
• Vaccine immunogenicity and reactogenicity runs all six analyses in the browser using the shared R numerical core. Upload mapped study CSVs or explicitly load the fictional example; no software installation or account is needed. Confirm assumptions and review missingness before interpretation. Download results before closing or reloading the session.

Validation & QC

• Trial-Management Validation Harness: ten gates (reproducibility, PK and count conventions, data structure and integrity, SAS/R twin structural parity, an adversarial QC panel, reporting conventions, AI-use documentation, package consistency, an independent recomputation of the sandbox results, and a seeded-defect test of what the gates catch). The checks run on this repo's own libraries and synthetic data and write an evidence ledger: 8 of 10 green, the QC panel is a manual record, and the AI-use record awaits sign-off. → validation_harness/

---

Principles the work is built on

• Validated tools own every reported number (Phoenix WinNonlin, Pinnacle 21, the EDC): a language model never produces a regulated number; AI drafts the words around a number, behind a human gate.
• Operations only: no PHI, no participant-level data and no unblinded assignments leave the client's systems.
• "Participant," not "subject" in prose; CDISC variables (USUBJID/SUBJID) preserved exactly.
• A named statistician signs every deliverable. Automation drafts and checks; it never decides.
• Run locally where confidentiality and reproducibility matter; a pinned, frozen model for anything regulated.

---

Notes

• Synthetic example data. The generator's design structure is informed by public early-phase registration patterns on ClinicalTrials.gov; all participant-level data is simulated from generic domain priors. Nothing here is real trial data.
• These are working templates and tools to adapt and validate for each client study, not validated production systems.

Design

The original pages (home, Operations, Biostatistics, the full library and the module pages) keep their layouts and use a shared light/dark adapter with site.css, site.js, _mc-topbar.html and _quarto-mc.scss. The newer pages (the worked study, the method studies, the harness page and the app launch pages) are built by build_pages.py from assets/theme/tokens.css into assets/ (toolkit.css, ui.js) and _partials/; python3 build_pages.py --check confirms they are current.

Hosting and package checks

GitHub Pages builds the site with GitHub Actions (.github/workflows/pages.yml): it copies the static files, exports the four R apps with Shinylive, and opens every app in headless Chrome before deploying; nothing is published unless every app passes.

Run python3 package_qc.py from a complete checkout to check navigation, local links, retired content, terminology and documented counts. GitHub Actions runs it, python3 build_pages.py --check and the worked study (study/run_all.R) on pushes and pull requests. These check the package, not the statistics.

Reusable launcher checks live in tools/tests/. Theme sources are assets/theme/tokens.css and assets/legacy-theme.css/js; run python3 build_pages.py after changing them or re-rendering a legacy page. The build inlines the theme into portable pages, and --check detects drift.

License

MIT; see LICENSE.