What 1,959 research papers say about missing healthcare data

Near the end of almost every clinical research paper is a paragraph few people read: the limitations. It's where authors say, in writing, what their study couldn't do, and usually why. Very often, the reason is data.

We built software to read those paragraphs at scale. Here's what 1,959 of them say.

What we looked at

Our Demand Intelligence Agent (DIA) searched Europe PMC for papers published between 2023 and 2026 that state a data limitation, across 11 disease areas, taking up to 200 recent papers per area. For each paper it records the exact sentence where the authors describe the gap, their organisation and the date, and sorts the gap into a type.

What researchers say they're missing

The share of papers stating each type of gap:

• Single-centre data: 47%. Almost half say their evidence comes from one hospital or site, so their results may not hold elsewhere.
• A larger cohort: 16%. Too few patients for the question they wanted to answer.
• External validation: 13%. A model or finding not yet tested on an independent group of patients.
• Under-represented populations: 12%. Groups the study couldn't include in sufficient numbers.

The pattern is clear. The thing researchers most often say they lack isn't a better method or more computing power. It's data from beyond their own hospital.

Why this matters

For research teams: if your study was single-centre, you're in good company, and reviewers will keep asking for the same thing. Planning multi-site or external validation data early saves a second study later.

For hospitals, labs and imaging centres: the data you hold is what these researchers say they need. Data from a different site, population or health system is what turns a single-centre result into evidence that holds. Our Data Collaboration Programme is how data owners make their data available for research, under governance, with a share of licensing revenue.

For companies building AI or diagnostics: the same gap applies to products. A model validated at one site hasn't yet met the rest of the world. See where public imaging datasets fall short for commercial AI.

Our own limitations

In the spirit of the paragraph we studied:

• We only see papers indexed in Europe PMC, mostly in English.
• Software sorts free-text sentences into types, and it won't be perfect.
• A stated need isn't a budget. Some authors will never look for more data; others need it but can't fund it.
• Papers report gaps months or years after the work was done.

Even so, a limitations section is one of the few places where people describe their data needs in their own words, with their name and organisation attached.

How DIA uses this

DIA turns these sentences into a list of organisations that have published a specific data need, with the evidence for each, alongside grants, clinical trials and regulatory filings. Data owners use it to see who needs data like theirs. See how DIA works.