Desktop
Early accessFor a programmer or data manager with one study on one secured PC.
- Runs on
- One Windows or Linux PC, local only
- AI
- Local model (e.g. via LM Studio)
- Licence
- Offline signed file, no call-home
EDC export → SDTM → ADaM → define.xml → eCTD m5
From raw study data to submission-ready SDTM, ADaM and Define-XML.
AI proposes each mapping with its reasons. Your data manager approves it. The conversion itself is deterministic, and every decision is signed and written to a hash-chained audit trail under 21 CFR Part 11.
Runs on one PC, on your own server, or (soon) in our cloud. Subject-level data never goes to an external AI model.
Mapping workbench
DEMO-001 · VS, DM, AE, CM, LB
SDTMIG v3.4CT 2026-09-25
5 AI suggestions
0 approved
0 need review
0 sent outside your network
| Source field | SDTM target | Confidence | Why | Status |
|---|---|---|---|---|
| VITALS.SYSBP“Systolic BP (mmHg)” | VS.VSTESTCD = SYSBP | ✓ High | Label and unit match codelist C66741 | ◇ Suggested |
| DEMOG.SEXvalues: M, F | DM.SEX | ✓ High | All values in non-extensible C66731 | ◇ Suggested |
| AE.SEVERvalues: 1, 2, 3 | AE.AESEV | ✓ High | CRF decode 1–3 → MILD / MODERATE / SEVERE (C66769) | ◇ Suggested |
| CONMED.DRUGNMlocal brand names | CM.CMTRT | ▲ Medium | Drug dictionary match; recode needs a signed review | ◇ Suggested |
| LAB.ALT_U“IU/L” | LB.LBORRESU | ▲ Medium | Keep collected unit; standard unit set in LBSTRESU | ◇ Suggested |
SIGN mapping v12 · reviewer · meaning “Reviewed against SDTMIG v3.4” · re-authenticated
AI_GATE task map_variables · declared L1 · found L1 · allowed L1 · sent
chain head 9f2c41…e1a0 · anchored
Built on
SDTMIG v3.4
ADaMIG v1.3 + ARM
Define-XML 2.1
NCI CDISC CT
CDISC CORE
XPT v5 · Dataset-JSON 1.1
Checks for
FDA · PMDA · EMA · MFDS
m5/datasets/<study>/tabulations · analysis
One study, start to submission package
Standards come from version-pinned CDISC metadata, not from a model's memory. AI drafts; rules check; people decide. The same approved specification always produces the same files.
- 01
Ingest
EDC exports, CSV, SAS7BDAT or XPT, plus the protocol. Every value is kept as text and profiled.
raw/ + SHA-256 - 02
Map
Domain and variable suggestions with confidence, reasons and the codelist each value was chosen from.
mapping specAI ledger - 03
Review and sign
Accept, edit or reject each suggestion. Recodes of free text are reviewed and e-signed. People only.
SIGN events - 04
Build SDTM
USUBJID, --SEQ, --DY, EPOCH, baseline flags, trial design domains and SE derived by rule.
*.xptDataset-JSON - 05
Derive ADaM
ADSL, BDS, OCCDS and TTE with traceability back to SDTM, and analysis results metadata.
adsl.xpt …ARM - 06
Describe
Define-XML 2.1 checked against the schema, with draft cSDRG and ADRG reviewer guides.
define.xmlcsdrgadrg - 07
Validate and package
CDISC CORE and ADaM checks, your Pinnacle 21 report imported, agency-specific submission checks, signed release.
eCTD m5 zipmanifest.sha256
90%
of SDTM cell values identical to the hand-programmed datasets delivered for the same real oncology study (anonymised). Most remaining differences are documented choices where we followed SDTMIG and CDISC CT over the original delivery.
100%
variable, Core and Controlled Terminology conformance across six public protocols when mappings are grounded in CDISC metadata, against 88% CT accuracy from a model's memory alone. That gap is why standards never come from the model.
Internal measurements, September–October 2026. Aggregates only; no study data leaves the sponsor's environment.
AI_GATE · every outbound call
Your subjects' data stays inside your walls
A local model is the default. An external model is used only when the licence, the installation and the study all allow it. Even then, each request is inspected before it leaves.
- Content is checked, not labels. The gate searches the actual prompt for this study's subject IDs, identifier formats and raw free-text values.
- Refusals are logged too. The audit trail records the task, levels, counts and a SHA-256 of what would have been sent. Never the values.
- AI never signs. A suggestion is a draft until a person approves it.
| Level | What it contains | External model |
|---|---|---|
| L0 | Public standards: CDISC IG and CT text, candidate lists | Allowed |
| L1 | Study metadata: dataset and field names, labels, spec cells, protocol text. No values. | If the study allows L1 |
| L2 | Values detached from subjects: codes, verbatim terms, comments | Admin opt-in per study |
| L3 | Subject-level records and direct identifiers | Never leaves |
21 CFR Part 11 · EU Annex 11
Built and checked like the regulated system it is
Your QA team gets evidence, not promises. Installation produces its own qualification records, and each release is signed and verified file by file.
§11.10(e)
Hash-chained audit trail
Append-only, with a reason on every change. The chain head is sealed with an external key held by QA, so a rewritten file is detected.
§11.50 · §11.200
Electronic signatures
Name, UTC time and meaning shown before signing. Password re-entry, bound to the hash of what was reviewed.
§11.10(d) · (g)
Roles and workspaces
Admin, data manager, reviewer and viewer. Sponsors are kept apart in separate workspaces.
GAMP 5
IQ, OQ and PQ at install
The installer verifies the signed release and writes qualification evidence your QA can approve.
§11.10(c)
Deterministic packages
The same state gives the same bytes. Files are re-hashed at signing; any later change voids the signature.
Annex 11 §7
Verified backup and restore
Backups are hashed per file, a broken chain is refused, and restore compares the chain head.
399automated tests passed in the 1.2.0 release verification
89user requirements traced to tests in the validation package
4agency profiles checked before packaging: FDA, PMDA, EMA, MFDS
One codebase · three deployments
Run it where your data is allowed to be
Private
Early accessFor CROs, sponsors and hospitals whose data may not leave their network.
- Runs on
- Your server or VPC, TLS, Docker
- AI
- Local model inside your network
- Teams
- Roles, sponsor workspaces, team audit trail
Cloud
WaitlistFor small biotechs that want to start without installing anything.
- Runs on
- Our hosted service
- Isolation
- Separate storage and audit trail per customer
- AI
- External model behind the data-level gate
Design partner programme · 10 places
Bring us a real study
We are opening early access with a small group of sponsors, CROs and academic teams. You run a study through MediLake Studio with our standards team beside you.
- Six months free on the Desktop or Private edition, including setup and qualification support.
- Direct line to the engineers who build the mapping engine and the validation package.
- What we ask: a short feedback call each week, and permission to publish an anonymised case study if it goes well.
Prefer email? contact@medilake.io
Questions from standards teams
Frequently asked
Does it replace Pinnacle 21?
No. MediLake Studio runs CDISC CORE conformance rules and its own ADaM checks, and it imports your Pinnacle 21 Community or Enterprise report so issues and explanations flow into the reviewer guides. Pinnacle 21 runs on your machine, so your data stays where it is.
Which standard versions does it produce?
SDTM with SDTMIG v3.4, ADaM with ADaMIG v1.3, Define-XML 2.1 and SAS XPORT v5, with Dataset-JSON 1.1 alongside. Each study pins its standards and Controlled Terminology version, and the pin is recorded with a hash so results can be reproduced later.
How accurate is the AI mapping?
The AI only proposes. Every code value is chosen from version-pinned CDISC metadata, checked by rules, and approved by a person before anything is built. On a real oncology study, 90% of SDTM cell values matched the hand-made delivery, and most of the rest were deliberate corrections. Design partners see their own study's acceptance rates in the product's AI ledger.
Can our QA validate it?
Yes. Installation verifies the signed release and generates IQ, OQ and PQ evidence. A validation package with traced user requirements and SOP drafts is provided for your computerised system validation.
Which agencies does the submission check cover?
FDA, PMDA, EMA and MFDS profiles, including PMDA Japanese datasets, FDA dataset splitting above 5 GB, and eCTD v4.0 study data metadata.
Who is behind MediLake Studio?
Clupea Inc., the team behind the MediLake clinical data platform (EDC, CTMS and eTMF), with CDISC advisory work since 2017 and the founders of the first CDISC Korea Summit. Teams in Korea and India.