In October 2019 I gave a talk at the CDISC US Interchange in San Diego called “Synthetic Data and Predictive Data Modeling: Benefits in Driving Standards and Quality Outcomes.” It opened with an observation I still stand behind: to many, test data is a necessary evil. Synthetic study data was a fairly quiet corner of the field back then. It isn't anymore — and since I've spent most of the years since building what that paper described, this seems like a good time to go back and reread it honestly: what held up, what didn't, and what I underestimated.
What I argued in 2019
The paper named the three ways teams actually got test data, and why each fails. Zombie data: records cobbled from past test cycles, dragged forward on the assumption that the old data model still fits the new one. Misfit data: data borrowed from a similar study, with the gaps hand-patched — and with real privacy and unblinding risk riding along. Piecemeal data: a few records mocked up per form to satisfy individual test cases, entered by hand.
Against those I set what I called the axes of integrity — the congruity that real data has because real subjects lived the protocol: cohorts enrolled in a shared window, visits in sequence with dates that agree, assessments landing inside their visit windows, histories that predate enrollment, durations that end after they start. Mocked-up data violates these axes constantly, and every violation is a downstream test you silently can no longer run. I asked the audience to borrow a statistical concept and apply it to their test data: does it have the power to support reliable conclusions about the quality of your study build? For most teams, honestly, the answer was no.
The constructive half of the paper laid out a four-phase method: acquire the study's design metadata; set up a synthetic study deliberately (a subject roster, per-subject visit inventories, dataset templates, terminology); then conduct the study — generate data chronologically, visit by visit, so that later data can react to earlier events; and finally process the output into whatever downstream shape is needed. And buried in the middle was the sentence I now think mattered most: in synthetic data generation, edit checks and rules are prescriptive, not proscriptive. A rule written to catch bad data can be reverse-engineered to define what should be created. Quality logic, run in the opposite direction, becomes a generator.
What the industry did in seven years
Three developments moved the ground under that talk — all in the same direction.
Conformance rules became executable. Before the 2019 paper I had spent years founding and leading the CDISC SDS Conformance Rules team, writing down rule by rule what “conformant” means. That effort evolved into CDISC CORE: an open-source engine anyone can run, executing the official rule set. The rules stopped being documentation and became software.
The protocol became machine-readable. USDM and the Digital Data Flow work turned study designs — arms, epochs, visit schedules, activities — into structured data. In 2019 the schedule of assessments was a table in a PDF that a person transcribed into metadata. A study design you can parse is a study design you can execute.
Generative AI happened. The part everyone knows. The consequence that matters for clinical data isn't chat — it's that coherent, realistic structure became learnable at scale.
Where synthetic study data is now
The method in the 2019 paper was real — the appendix was working SAS code driving generation from a metadata sheet. But it was a craftsman's method: the supplemental metadata was authored by hand, and the rules were mine. What I have built in the years since — Verdatic — is that method grown into a platform, with each hand-tooled step replaced by something systematic.
From a study design and the standards themselves, it simulates a realistic multi-domain SDTM study: demographics, vitals, labs, ECGs, medical history, adverse events, and concomitant medications, on visit schedules derived from the machine-readable design, with treatment arms that behave like treatment arms — a modeled drug effect, not noise. The 2019 “axes of integrity” are no longer a checklist to aspire to; they are enforced by construction, because subjects are generated as coherent digital twins who experience the protocol chronologically — exactly the “synthetic study conduct” idea, industrialized.
And the prescriptive-rules sentence became the heart of the system. The obvious use of CDISC CORE is checking data after the fact. The better use is closing the loop: generate, run the official engine, feed the findings back into generation, iterate until the verdict is clean. Our working corpus now runs the official CORE engine to zero open data defects. The rule set I helped author to catch problems now prevents them.
The downstream half of the 2019 argument — what the paper called the synthetic data “waterfall” — is where the productization shows most. Synthetic data that exists before first patient in lets you rehearse the entire pipeline: mapping specifications, value-level metadata, and generated SAS, R, and Python transformation programs, through SDTM to ADaM. In 2019 I described a sponsor producing synthetic ADaM before their EDC system was in production, as an outer limit of what disciplined metadata makes possible. Today that sequence is a feature. Likewise the UAT story: the paper's coverage diagrams showed test subjects scattered thinly across a schedule of assessments; now paired clean and deliberately dirty twins are generated from a study's own edit checks — PASS cases and FAIL cases, graded automatically.
What held up, and what I underestimated
The benefits case held up in full — I'd argue it was conservative. Test data with real power does accelerate study builds, qualification, and downstream program development, and the 2019 use case I flagged for the standards community — using synthetic data to evaluate a new standard's impact before adoption — has only grown more relevant as standards versions turn over faster.
What I underestimated was my own best idea. In 2019, “prescriptive rules” meant a study's edit checks, hand-translated into generation logic by someone like me. I did not foresee how far the inversion would travel: the entire official conformance rule set, executed by an open engine, teaching a generator in a self-correcting loop — with people supervising the loop rather than hand-feeding it. When the simulator can't satisfy a rule, sometimes the data is wrong, and sometimes the rule is. Both discoveries are valuable, and both now happen before a single patient is enrolled.
One belief unchanged from 2019: none of this removes people. Synthetic data is for development, testing, training, and demonstration — not regulatory submission — and the judgment calls stay human. Agents propose; people decide.
Where this goes next
This year Verdatic is an entrant in the CDISC AI Innovation Challenge, and its self-correcting approach to simulated study data was selected for presentation at the 2026 CDISC Interchange — the same stage as the 2019 paper, one wave later. In 2019 I brought an argument. This time I'm bringing the working system that came out of it, and I'm looking forward to the conversation.
If you want the current state of the argument, the demo is at verdatic.com/challenge. If you want to talk about what synthetic data could do for your studies, that conversation is what Clinventive is for.