biomarker-sdtm

biomarker-sdtm

An agent skill for transforming clinical data. It converts documents produced from a study into CDISC SDTM outputs, and generates reproducible computational records that follow the official BioCompute Object standard (IEEE 2791) alongside a BiomarkerKB record built on a unified biomarker knowledge framework.

🔬 Generate submission-ready SDTM from your study documents

Provide the documents a biomarker study already produces: assay specification, context of use, method validation report, patient results; and the skill returns CDISC SDTM domains ready for downstream review:

DM · LB · SUPPLB · BM + further domains required

BM is a proposed SDTM domain for Biomarker data. It supplies the biomarker context for an assessment that already exists elsewhere, rather than duplicating patient-level results.

This skill also supports the reverse direction: given existing SDTM output, it reconstructs the corresponding BioCompute Object and BiomarkerKB records, so legacy tabulations can be brought into the modern pipeline with ease.

flowchart LR
    docs["Study Documents<br/>(protocol · context · assay specs · results)"]
    agent["Agent"]
    sdtm["CDISC SDTM<br/>(DM · LB · SUPPLB · BM ...)"]

    subgraph comp["Biomarker record set"]
        direction TB
        tbco["¹ tBCO<br/>assay specification"]
        rbco["² rBCO<br/>per-patient run"]
        bmkb["³ BMKB<br/>BiomarkerKB record"]
    end

    docs --> agent
    agent --> tbco
    agent --> rbco
    agent --> bmkb
    tbco --> sdtm
    rbco --> sdtm
    bmkb --> sdtm
    sdtm -. "reverse" .-> agent

Each full run produces a biomarker record set that travel with the data:

¹ tBCO: the template BioCompute Object, the assay specification

² rBCO: the run BioCompute Object, one record per subject per analytical run

³ BMKB: the BiomarkerKB record, what the biomarker means in controlled vocabulary

SDTM records what the measured values were. The record set records how they were produced and what they mean, so the study remains reproducible and comparable long after the spreadsheet is gone.

Standards & knowledge frameworks

BioCompute Object (BCO)

A BCO is a machine-readable record of how a computation was performed: the scientific intent, the computational steps, the parameters, the software, and the inputs and outputs. It is an IEEE standard (IEEE 2791), community-driven and peer-reviewed. For clinical research it supplies a regulatory-grade audit trail, allowing an analysis to be re-executed and verified years later rather than accepted on trust.

BiomarkerKB

BiomarkerKB is a knowledgebase that harmonizes and standardizes biomarker data. It captures the measured entity, the condition, and its role (diagnostic, prognostic, monitoring, and others) in a controlled vocabulary, so that a finding from one study is directly comparable with a finding from another. Full project description can be found in Patterns (2026). For clinical research, it converts one-off biomarker measurements into structured evidence that can be pooled, queried, and reused across trials.

The two close the gap that SDTM leaves open: SDTM standardises the tabulation, while BioCompute Objects and BiomarkerKB standardise the provenance and the meaning.

Both projects are developed at George Washington University. BioCompute was developed with support from the U.S. Food and Drug Administration (-2026). BiomarkerKB is funded by the NIH Common Fund Data Ecosystem (CFDE).

Resources

• BioCompute Object Portal: https://www.biocomputeobject.org/
• IEEE 2791-2020 (BCO standard): https://standards.ieee.org/standard/2791-2020.html
• BiomarkerKB Portal: https://biomarkerkb.org
• BiomarkerKB data submission: https://wiki.biomarkerkb.org/Data_Submission/Data_Upload
• CDISC SDTM: https://www.cdisc.org/standards/foundational/sdtm
• BEST (Biomarkers, EndpointS, and other Tools) Resource, FDA-NIH Biomarker Working Group: https://www.ncbi.nlm.nih.gov/books/NBK326791/

Quick Start

[!Note] Skill uploads are not supported on mobile devices. However, you can upload them from a computer and then use them on your phone while signed into your account.

1. Download the repository as a ZIP

• On the repository page, choose < > Code → Download ZIP.

2. Upload the ZIP to your preferred application

Claude

• Go to Settings → (Side bar) Customize: Skills → + Add dropdown → Upload skill.
• Drag and drop the ZIP file to the drop zone → Upload.

ChatGPT

• Go to Plugins → Toggle to Skills → Add → Upload from your computer.
• Drag and drop the ZIP file to the drop zone to upload.

Gemini

• Go to Settings → Preferences: Skills → Create dropdown → Upload.
• Drag and drop the ZIP file to the drop zone to upload.

3. Invoke it from any chat with the slash command:

• Upload your clinical documents using + (Add files).
• Invoke the skill with the slash command:

/biomarker-sdtm

We have included a set of simulated clinical documents in the skill package for demos. Try:

/biomarker-sdtm help me build a full biomarker record set and SDTM output from the troponin I study documents under the skill's fixtures/ folder. You are not allowed to open anything under the example/ folder in the process.

[!TIP] Claude also loads SKILL.md automatically when a request matches it. Mention BioCompute, a tBCO or rBCO, BiomarkerKB, or SDTM and it will pick the skill up without the slash command.

Layout

biomarker-sdtm/
├── SKILL.md              # workflow (start here)
├── README.md
├── CHANGELOG.md
├── references/           # docs + schemas, loaded on demand
│   ├── README.md         # index, naming and versioning convention
│   ├── bmkb-model.md     # BiomarkerKB submission schema + accession lookup
│   ├── sdtmig-guide.md   # SDTMIG rules for DM, LB, SUPPLB, BM
│   ├── sdtmig-v3.4.pdf   # the full SDTMIG v3.4 specification
│   ├── bco-schema.md     # IEEE 2791 field spec
│   ├── best-definitions.md
│   ├── pitfalls.md
│   └── schemas/          # IEEE 2791 JSON Schemas
├── scripts/              # accessed by agent at runtime
│   ├── finalize_draft.py        # strip //_ annotations, compute/verify the etag
│   ├── validate_bco.py          # IEEE 2791 structural validation
│   ├── validate_bmkb.py         # BMKB submission-schema validation
│   ├── biomarker_accessions.py  # entity type → accession resource
│   ├── censor_state.py          # analytic censoring logic
│   ├── record_set_to_sdtm.py     # record set → SDTM
│   └── sdtm_to_record_set.py     # SDTM → record set
├── assets/              # annotated templates the agent fills
├── examples/            # VALIDATED reference outputs + adjudication log
└── fixtures/            # SIMULATED input documents for testing

Compatibility checks (Developers)

Checks to confirm the helpers run correctly in your environment. No documents required.

python scripts/run_checks.py

It runs every script's self-test plus a comparison against the Troponin I gold example, and exits non-zero if anything fails.

To score an agent's test run (its output folder) against the gold, on SDTM cells and controlled record-set fields only:

python scripts/grade_run.py <run_folder>

Standard versions

This work is supported by BiomarkerKB.

Disclaimer

[!IMPORTANT] This skill is a research and development tool. It drafts structured records from study documents; it does not make decisions, and it does not submit anything to any regulator, sponsor or data provider. Everything it produces is a draft requiring expert review before it is used for anything.

• Not a regulatory submission. The SDTM domains, BioCompute Objects and BiomarkerKB records it produces are drafts. They are not validated for submission to any health authority, and this skill does not provide regulatory advice. The sponsoring organisation remains responsible for its own submissions and for verifying every value against the source documents.
• Verify before relying on it. Automated extraction from documents can be wrong, and a PLACEHOLDER marks a value the documents did not supply; it is not a default. Treat every generated field as a proposal to check, not a fact.
• Standards and schemas move. The output formats follow the versions listed above. Confirm them against the current published versions before submitting.
• No warranty or liability. The skill is provided "as is", without warranty of any kind, express or implied, including fitness for a particular purpose. The authors and their institutions accept no liability for any loss or damage arising from its use.