What Real-World Data Sources Actually Look Like

Every source of real-world data was built to do a job, and almost none of those jobs was research. A registry tracks a defined group of patients, a health record cares for one person in front of a clinician, a claims file pays a bill, and a wearable sells a consumer a habit. Draw a medical conclusion from any of them and you are borrowing a tool built for something else, its original purpose stamped into what you can honestly say. The first question about any real-world finding is not how large the dataset was. It is where the numbers came from and what they were made to do.

I use real-world data to mean information generated during ordinary life and care rather than inside a designed experiment. The four sources below cover most of what a modern study runs on. Each captures a real slice of the world, and each leaves out a different slice, so the same clinical question can look answered in one source and unanswerable in another.

What do patient registries capture well, and what do they miss?

A registry is a curated list of people who share a condition, an exposure, or a procedure, followed forward on purpose. Because someone designed it to answer questions, it tends to collect the variables that matter and to define them consistently, which is exactly what most other sources fail to do. A well-run disease registry can tell you how a defined population fares over years, including the slow outcomes a short trial never reaches.

The weakness of a registry is who gets into it and who keeps showing up. Enrollment is rarely a random sample of everyone with the condition, so a registry can quietly over-represent patients seen at large centers or those healthy enough to keep returning for follow-up. When people drop out, the ones who remain often differ from the ones who left, and the registry keeps recording the survivors as if they spoke for all. A registry is only as representative as its front door and its retention.

Why are electronic health records rich but treacherous?

An electronic health record holds the deepest clinical detail of any routine source: lab values, notes, diagnoses, imaging, the actual numbers a clinician acted on. For questions that need physiology rather than billing categories, nothing else in ordinary care comes close. This depth is why records have become the backbone of so much observational research and of the clinical AI trained on it.

The trouble is that a record documents care, not truth, and only the care that happened inside one system. A patient who sees clinicians across three unconnected networks appears in each as a fragment, healthier than reality because half the history is missing. A test that was never ordered looks identical to a test that came back normal, since both are simply absent. Sicker patients are measured more often, so the frequency of a lab draw carries hidden information about how worried someone was. None of that is error. It is the residue of real clinical decisions, and it will fool any analysis that reads a blank cell as a normal value.

What can insurance claims tell you, and what can they never see?

Claims data is generated to pay for care, and that gives it two real strengths. It is comprehensive across settings, because everything billed under one plan flows into one stream regardless of which hospital or pharmacy delivered it, stitching together a fuller path than a single record system. It is also excellent at capturing events that reliably generate a bill: a hospitalization, a procedure, a filled prescription.

What claims cannot see is anything nobody charged for. There is no lab result in a claim, only the code that a test occurred. A diagnosis code may reflect what was billed most efficiently rather than what the patient actually had, because coding answers to reimbursement rules before clinical precision. A filled prescription tells you the medicine left the pharmacy, not that anyone swallowed it. So claims are strong for counting expensive, discrete events and weak for severity, physiology, or whether a treatment was truly taken. The conclusion the data supports is about billed events, honest only when the clinical question lines up with what generates a bill.

Why are wearable and consumer-device data both promising and slippery?

Wearables and consumer sensors offer something no clinic can: dense, continuous measurement in the wild, thousands of readings from someone living an ordinary life instead of a handful captured during a visit. For patterns that only show themselves over time, such as day-to-day variability or a slow drift, this stream can reveal what a periodic measurement would never catch.

The catch begins with who owns a wearable. Device users skew younger, wealthier, and more health-motivated than the general population, so any group defined by having the device is already selected before a single number is analyzed. The measurements themselves vary in accuracy across devices and are usually not validated to clinical standards. People also wear and abandon devices in ways that track their motivation, so gaps in the data are informative rather than random, and a burst of readings may mark a good week rather than a typical one. Rich sampling of a self-selected, loosely calibrated population is a genuine asset, but only for questions that survive those facts.

How should the source change the conclusion you draw?

Match the claim to what the source was built to observe, and distrust any conclusion that leans on the part the source cannot see. Ask what job the data was originally doing, who entered it and why, who is missing, and whether a blank means normal or unrecorded. The same signal deserves more trust when it appears across sources whose weaknesses do not overlap, because a bias that inflates a registry rarely inflates a claims file in the same direction.

In building clinical decision-support tools, I lived this distinction daily. Usage logs and records were invaluable for understanding how a system behaved in real clinics and where a workflow broke. But for the central claim, whether the tool changed patient outcomes, we relied on a registered randomized controlled trial (NCT03258268), because no routine source was built to answer that question cleanly. The peer-reviewed work I trust most, including a meta-analysis of ethnic differences in insulin response, earned its weight by being explicit about where its numbers came from. That habit is the whole discipline. This article is general education, not medical advice, and any decision about your own care belongs with a qualified clinician.