Every trial result rests on biometrics: the clinical data management, biostatistics and statistical programming work that turns raw trial data into decision-ready evidence. That work is changing faster than it has in decades. AI can now draft code, flag outliers and map data to standards. ICH E6(R3) asks teams to collect less and justify more. And wearables are producing data that no case report form was built to hold.
The question is no longer whether these changes will reach clinical data management. It’s whether teams can adopt them without giving up the rigor regulators and patients depend on. The same question ran through the program at the Society for Clinical Data Management (SCDM) 2026 Annual Conference in Raleigh this month, where we demoed our intelligent submission software, Lodestar™.
AI in Clinical Data Management: Ready, but Can We Trust It?
AI is ready to help with clinical data, but teams will only rely on it once they can show what it was validated to do and who is accountable for its output. Trust is a process question, not a model question.
One session SCDM named the problem in its title: “Why AI Pilots Don’t Make It to Production.” Its abstract, from chair Sina Adibi of Adaptive Clinical Systems, is blunt: “The issue is usually not the model or the math.” Clinical data work runs on judgment and review, and “trust is built over time.”
In the opening keynote, Michael Pencina, Chief AI Scientist at UnitedHealth Group, described AI “moving from pilot projects to embedded infrastructure,” arguing that responsible AI “can’t be a slogan” and has to be “a framework, built on the same rigor clinical data managers already bring to source data and study design.”
Other sessions pressed on what that rigor looks like. One panel asked how well a sponsor needs to understand a vendor’s AI model to validate its intended use.2 Another examined “how often AI assisted humans defer to wrong suggestions,” a pattern known as automation bias: the habit of accepting a computer’s answer even when it’s wrong.
The U.S. Food and Drug Administration (FDA) points the same way. Its January 2025 draft guidance asks sponsors to define a model’s context of use, assess its risk, plan how its credibility will be established, and maintain that credibility over the model’s life cycle.
In practice, this means four things:
- Settle the context of use before a tool reaches study data
- Validate against that purpose.
- Keep validating as the model or study changes.
- Make a named person responsible for every AI-assisted decision.
We built Lodestar™. around that principle. “It will make decisions. However, it always gives the opportunity for the human review and acceptance of those decisions,” says Michael Mendoza, Koneksa’s SVP of Biometrics and Data Science. “You can override, and as you override each thing, it’s going to require that rationale.” Independent biometrics advisor Shari Medendorp of SVM Life Sciences puts it simply: “The most important thing as we move through new technology is that there is a human and expert in the loop.”
ICH E6(R3) Wants Leaner Trials. There Catch is Timing.
ICH E6(R3) lets teams stop collecting and checking data that doesn’t matter, but the savings only hold if those decisions are made and documented early.
In the weeks before database lock, queries are everywhere, and every field is checked, then checked again. Teams often chase “error-free” datasets so that nothing can be questioned later. Much of that effort goes to data that rarely changes a trial’s conclusions.
The final ICH E6(R3) guideline was adopted in January 2025,4 and FDA published it as final guidance in September 2025. At its heart is proportionality: matching effort to risk. Principle 7 says trial processes should be “proportionate to the risks to participants and to the importance of the data collected.” Annex 1 goes further, telling sponsors to “avoid unnecessary complexity, procedures and data collection” and not to “place unnecessary burden on participants and investigators.”
That theme ran through SCDM. A panel titled “The ICH E6(R3) ‘Stop List’: Practical Proportionality” proposed legacy tasks E6(R3) empowers teams to retire. It described teams that “remain paralyzed by the legacy of ‘error-free’ data” and set out how to document the rationale for what isn’t checked. A session marking the guideline’s first year argued that good data “is no longer defined by how clean the database looks at lock,” but by how trustworthy and reconstructable it is across its full lifecycle. And the closing panel called for research that is “more proportionate, practical, and patient-focused,” with less unnecessary data collection and sites involved earlier in study design.
Here’s the catch. Every decision to stop is a decision you have to justify. Make it late, and you drop data you’ve already paid for, with a rationale that’s harder to defend.
The fix is to decide what to collect while the protocol is still being written, when the reasoning is fresh and easy to record. Kimberly Guedes, VP of Clinical Operations at Intensity Therapeutics, knows the cost of finding out late: “I want to know in the beginning before I have to do any shifts.” Once agreed, a stop list becomes part of the measurement strategy, and the trial gets leaner by design.
Your Trial Data Has Outgrown the Case Report Form
Continuous and event-driven data can’t be retrofitted into a case report form, so its structure has to be designed up front, alongside the measurement strategy.
The case report form (CRF) records the information a protocol requires for each participant. It assumes the team knew in advance exactly what each field would hold. That assumption is breaking. At SCDM’s regulatory town hall, organizers noted that wearables, ePRO (electronic patient-reported outcome) platforms and real-world data sources “are part of the standard toolkit for clinical trials.”
A wearable streaming data around the clock doesn’t answer a fixed set of questions at a visit. It records a participant’s daily life, and meaning emerges from patterns over time rather than any single value. That’s the promise of digital biomarkers, and it’s why one SCDM session traced the shift “from established laboratory assays to high-dimensional and digital biomarkers,” data sources that “extend far beyond traditional case report forms.”
Event-driven trials stretch the model too. Their statistical power comes from clinical events accumulating over time, not fixed visit schedules, so the details around each event matter. One session, “Establishing Scalable Data Standards for Event-Driven Outcomes Trials,” chaired by Andrea Milner of Eli Lilly, noted the industry still lacks consistent standards for elements such as survival status, censoring and end of follow-up. Its proposed fix: treat standardization “not as a downstream cleanup exercise but an upstream design discipline.”
So the meaning of “clean” is shifting. Speakers in one SCDM session, working separately, reached the same conclusion: “clean,” “reviewed” and “locked” have “quietly stopped meaning ‘analysis ready.'” Take a heart rate reading. It means little unless you know when it was taken, whether the device was being worn and which software version produced it. FDA guidance on digital health technologies asks sponsors to capture that kind of metadata and to record device updates.6 Clean now means the signal can be interpreted.
The answer is measurement strategy first. Decide what needs to be measured, and how often, before you choose how to capture it. Then set the data structure and standards to match. Pick the device first, and your study inherits whatever shape of data that device produces. The profession is already adapting, with the data manager’s role “expanding beyond data review and query management.”
From Database Lock to Submission: Where AI Is Already Delivering
The stretch between database lock and regulatory submission is where fragmented tools and manual handoffs cost programs the most time, and where connected, AI-assisted workflows are already showing results.
Discovery has accelerated over the past decade. Submission has not. Validation, metadata, programming and eCTD packaging still tend to live in separate tools owned by separate teams, stitched together by hand. “Every piece of the submission stack existed, but nothing connected it,” says Mendoza.
That’s why we built Lodestar™, the first intelligent submission software. It connects biometrics, QC and eCTD assembly in one AI-assisted, audit-trailed workflow, with expert approval at every stage gate. When we ran our own Parkinson’s disease study through it, Lodestar™ delivered:
- 80% reduction in programming hours
- 65% lower biometrics costs
- 16x faster QC cycles
- 100% audit-ready outputs
We demoed Lodestar™ at booth 244 at SCDM 2026. If you missed us in Raleigh, you can request a demo or watch the on-demand webinar, Meet Lodestar™: Your North Star for Regulatory Submissions, with Mendoza, Guedes, and Medendorp.
Looking Ahead
The field isn’t short of data. It’s working out which data it can stand behind, and that’s a good problem to have. The teams that benefit most from AI, E6(R3) and continuous data will treat validation, proportionality and data structure as design decisions, not cleanup. Decide what to measure, and why, before the first participant enrolls. Then the data that follows can become some of the clearest evidence a trial produces.
If you’re planning endpoints or a submission for an upcoming study, request a Lodestar™ demo or connect with our biometrics team today.
