AI Is Hungry For Data: Does Cross Country Data Make Trials Better?

By Deepti Kalghatgi, chief curator, IDAAHub

Every AI model is only as good as what it was trained on, and clinical trials are where that truth gets tested. The two systems want the same thing — more data — but they want it differently. AI wants volume. Trials need range: enough variation across age, sex, ancestry, geography, and comorbidity that a result holds up for the people who'll actually take the drug. That difference doesn't stay contained to study design. It shows up on the loading dock, in the resupply algorithm, in how much drug gets shipped where — because every demographic gap a trial has to fill becomes a site, a country, or a population segment that the supply chain didn't originally plan around.

The Historical Record On What Happens When Trials Lean One Way

This isn't theoretical. Cardiovascular disease interventions were developed on evidence drawn mostly from male participants, despite cardiovascular disease being the leading killer of women in the U.S. and presenting differently in their bodies. Common asthma treatments are documented to be less effective specifically for people of African American or Puerto Rican heritage, a gap that traces straight back to who wasn't well represented when those drugs were tested. In oncology, Black participants accounted for just 2.9% of drug company-sponsored cancer trials in a 2020 analysis, even though representation gaps like this directly limit the ability to catch early safety signals for the patients most affected. Even Moderna's COVID-19 vaccine trial hit this in real time: when reporting showed Black Americans made up only 7% of enrollment against 13% of the U.S. population, the company made the unusual call to slow its own trial down specifically to recruit more participants of color before moving forward. Every one of those mid-course corrections is a supply chain event in disguise — a sudden push to enroll a specific population, at specific sites, that nobody forecasted at trial launch.

Does Cross Country Data Actually Make Trials Better Or Just More Complicated To Supply?

The research says better, unambiguously — but with a catch that matters operationally: diversity has to be the right kind, not just more countries on a map. A 2025 analysis found that trials supporting recent FDA approvals did include larger proportions of non-white participants than a decade earlier — but only one in four of those participants actually lived in the United States, because sponsors increasingly run trials overseas for cost reasons. That's a real question for supply planning, not just an ethics one: a trial spread across more countries to capture demographic range means investigational product has to move, clear customs, and stay within stability windows across more borders, more cold chain handoffs, and more regulatory regimes than a single-country trial ever did.

This Is Exactly Where AI Raises The Stakes Instead Of Lowering Them

A 2026 review of FDA's draft AI guidance makes the connection explicit: AI models trained on narrow data don't just produce bad recommendations, they produce confidently bad ones, and regulators are now requiring sponsors to validate AI performance separately across demographic subgroups rather than trusting one blended accuracy score. The FDA's own 2024 draft guidance on multiregional oncology trials names the specific failure mode statisticians call covariate shift: when a trial's overall demographic mix drifts far enough from a target country's population that regulators flag the data as potentially not applicable there at all, even though the trial “succeeded” globally. When that happens, the fix isn't just a protocol amendment — it's adding sites, often quickly, in the specific country or population segment that's underrepresented. That's a demand spike with almost no lead time, landing on a supply chain that built its forecast around the original site list.

Where This Lands On Supply Chain Specifically

• Demand forecasting breaks at exactly the moment diversity gets corrected. RTSM/IRT systems forecast off historical enrollment pace. A late-stage push to fix a covariate shift gap, or a Moderna-style deliberate slowdown-then-acceleration in a specific demographic, doesn't look like anything in the historical data the forecast was trained on.
• More countries for diversity means more cross-border handoffs for drugs. Each additional country brings its own import licensing, customs clearance, and cold chain requirements — multiplying the points where a shipment can sit, expire, or get held, even before AI-driven enrollment speed changes how fast that drug needs to arrive.
• AI-driven patient matching tools tuned for diversity targets will enroll unevenly by design — fast at some sites, deliberately slower at others until subgroup targets are hit. A supply plan built on smooth, even ramp-up will consistently over-supply some sites and under-supply others.

The Challenge Nobody's Fully Solved Yet

The FDA wants enough domestic representation to trust a result for U.S. patients specifically; AI needs broad cross-population data to avoid the bias failures documented above; and sponsors need multiple countries involved to enroll fast enough to finish the trial at all. Those three needs pull in compatible directions in theory, but in practice, every adjustment made to satisfy one of them shows up as an unplanned shipment, a rerouted batch, or a site that suddenly needs drug it wasn't allocated. The diversity question used to belong to clinical operations. Increasingly, it belongs to the supply chain, too.

About The Author:

Deepti Kalghatgi is chief curator and runs an AI in healthcare hub called IDAAHub, an AI discovery and growth platform. IDAAHub curates specialized AI solutions and diverse data for healthcare organizations, helping them discover, evaluate, and adopt innovative AI solutions and diverse data for their AI models.

She previously founded ApptoHealth, an AI startup focused on care coordination, prior to COVID. Deepti is also an angel investor in various healthcare-based startups and is passionate about helping startups grow.

Her portfolio experience includes work with Eka Care, a company that has built one of the largest longitudinal patient data sets in India. This experience has informed her perspective on healthcare data systems and interoperability, which contributes to the thinking behind this article.