How AI Is Used To Predict Trial Performance Before Enrollment
By John Oncea, Chief Editor, Clinical Tech Leader
Most conversations about AI in clinical research still center on automation: faster data entry, quicker document review, less administrative burden. Nick Spittal, Chief Operations Officer of Velocity Clinical Research, and Raghu Punnamraju, the company’s Chief Technology Officer, are focused on a different application entirely: using AI and operational data to predict how a study will perform before it starts.
“There’s so much data available, and there’s an opportunity – and we spent quite a bit, and I’m sure we can do a lot more – in converting the data into knowledge,” Punnamraju said. For Velocity, that shift from data to what he calls “accessible knowledge” is the foundation for a set of predictive tools already reshaping how the company selects sites and plans operations.
It’s a distinction worth considering. Most site networks and CROs already collect enormous amounts of operational data: enrollment logs, staffing records, protocol histories. Far fewer have done the work to make that data queryable in a way that actually changes a decision before it’s made. Velocity’s leadership sees that gap, not raw data volume, as the real constraint on AI’s usefulness in clinical operations.
Rethinking Site Feasibility
One of the clearest examples is Velocity’s AI-supported feasibility process. Traditionally, feasibility assessments lean heavily on an investigator’s judgment and prior experience in a given therapeutic area – essentially, “I think I’ve got 10 of those patients in my database, and I can enroll three of them,” as Spittal put it.
Velocity’s approach goes deeper. Using large language models to parse incoming study requirements, the company breaks new protocols down to the underlying condition set rather than stopping at a therapeutic-area label. That reframing has surfaced sites that wouldn’t look like obvious candidates on paper but perform strongly once the comparison is made on patient population and operational history instead.
“If you look at it closer, the same condition set – site B has done the same level of enrollment rates compared to site A, just that it’s not part of the same bucket in the same therapeutic area,” Punnamraju said. Spittal called it “very data-driven,” pairing what a site’s team knows about its own patient population with a quantified record of past performance, something the industry has long treated as a qualitative judgment call rather than a measurable input.
From Reactive To Predictive Operations
That same data foundation is now being applied to workforce and capacity planning. Velocity holds years of historical data on enrollment pacing, staffing levels, and site activation performance; information that, until recently, required going through “line by line, study by study, site by site” to make sense of, Spittal said.
The goal is to automate that process into something closer to forward-looking staffing guidance: instead of a coordinator manually flagging a hiring need, the system would surface it months in advance, down to when recruiting should start and when a new hire needs to be in place, based on the study’s protocol, location, and required therapeutic experience.
That same predictive logic has already paid off in participant engagement. Velocity deployed engagement technology with an outside partner and has seen no-show rates improve by six to nine percentage points, according to Punnamraju, a result he described as directly applicable industry-wide, not something Velocity is interested in keeping proprietary. "Although we want to remain unique and differentiated, this is something for which the entire industry could benefit,” he said.
Enter The Digital Twin
Digital twins are a concept borrowed from manufacturing and aerospace, where engineers build a virtual replica of a physical system – a jet engine, a factory line – to test how it will behave under different conditions before touching the real thing. Velocity is applying the same logic to trial operations. The most forward-looking piece of this work is a digital twin concept the company has piloted for a small number of studies, with a fuller build targeted for later this year or early next.
The idea, as Punnamraju described it, is to use a study’s protocol, site history, and participant pool data to model likely recruitment performance before the trial launches, and then to test interventions against that model rather than the live study. “You use your data to forecast or predict how your study is going to do,” he said. “You also have the levers to play with in a simulated mode to see what you do right now to have a meaningful impact.” Instead of reacting to a slow enrollment curve months into a trial, the team could intervene at the design stage of a recruitment or engagement strategy.
Velocity is validating this concept the way any predictive model should be tested: “We withhold 20% of the data and have the model predict, ‘Hey, what would happen for this remaining 20%?’ And we see how aligned it is,” Punnamraju explained. Only after that pressure test would a model move into production, with a human expert still signing off. None of this is possible without groundwork Velocity put in years earlier: cleaning and standardizing its data before asking any model to learn from it. That same data discipline anchors the company's approach to AI governance more broadly, discussed below.
Prediction, Not Replacement
Neither executive frames this work as AI replacing human judgment. Spittal was clear that the value is in giving people better information earlier – surfacing a staffing gap before it becomes urgent or flagging an enrollment risk before it becomes a missed timeline.
That distinction matters for an industry where delays are often the product of information arriving too late to act on. Sponsors and CROs evaluating site partners increasingly ask for evidence of speed and enrollment performance up front; Spittal noted that showing "real data" behind those claims, rather than an anecdotal track record, has become table stakes with the sponsors Velocity works with, not a differentiator.
For sponsors specifically, the implication of a working digital twin extends beyond any one site relationship. A model that can forecast enrollment pacing before a study starts is also a model that could inform how conservatively a sponsor plans backup sites and contingency timelines across an entire trial.
If Velocity’s early results hold up, the payoff isn’t a fully automated trial; it’s a research organization that sees its own future clearly enough to change it before problems happen. That level of trust in a model doesn't come free, however. It depends on the same governance discipline that determines which AI tools Velocity is willing to put in front of a clinical or financial decision in the first place.