When do clinical records become research evidence?

An oncology record can contain a treatment date, a scan interpretation, a physician's judgment, and a note about why a plan changed. Those details are valuable, but they do not arrive as a ready-made research dataset. Flatiron Health built around the gap between documentation used to care for an individual and evidence used to study a population. The case is less about discovering that data exist than about the work required to make them reliable for a defined question.

In 2016, Flatiron announced a research collaboration with the FDA focused on information from patients with advanced non-small-cell lung cancer receiving immunotherapy. The stated purpose was to explore treatment patterns and the potential use of real-world evidence. FDA's own subsequent description of its oncology data work also identified the collaboration. Neither announcement meant that electronic records could simply replace randomized trials. The bounded move was to investigate a specific clinical setting and learn which questions the available information could credibly support.

Roche's 2018 acquisition announcement described Flatiron's oncology electronic-record software and its curation of real-world evidence. These two activities connect, but they serve different users. A clinician needs a workflow that supports patient care. A researcher needs consistent definitions, appropriate populations, and a defensible analytical dataset. A pharmaceutical buyer may value evidence relevant to development or use of therapies. The system has to reconcile those purposes without assuming that a record collected for one reason automatically satisfies every later purpose.

A peer-reviewed study of real-world progression across several cancer types illustrates the difficulty. The researchers described structured and unstructured information curated through technology-enabled abstraction and tested a progression variable across disease settings. Their methods also noted that care occurring elsewhere and not documented in the source record was unavailable. The study included company-affiliated researchers, so publication should not be confused with independence. It is useful because it exposes the definition and validation work behind an apparently simple endpoint.

Consider the practical difference between a scan date and a progression date. The first may be straightforward to locate. The second can depend on clinical interpretation, documentation conventions, and the research definition being applied. Missing information is not automatically evidence that nothing happened. A dataset can be large and still have systematic gaps. These are methodological implications of the source material, not allegations that a particular dataset is unusable. They explain why the provenance of a variable matters as much as the number of records.

Independent reporting on the extension of the FDA collaboration in 2019 highlighted continued work on data quality and validation of clinical endpoints. That continuation is a more disciplined signal than treating an acquisition price as proof of scientific usefulness. Real-world data can illuminate patients and treatment patterns that trials do not fully capture. They can also contain selection effects, incomplete follow-up, and confounding. The appropriate claim depends on the question, the methods, and the limitations, not merely on whether the source is called real world.

For a founder or research leader, the first move suggested by Flatiron is to choose one consequential question and trace every required variable back to how it was created. Who recorded it? For what purpose? What is missing, and who is absent from the dataset? What validation would make the proposed inference credible? Flatiron's story makes the connecting work visible. The valuable product is not just access to information. It is a dependable route from clinical practice to a research question, with enough discipline to show where that route ends.

For a discussion tailored to your team, explore a workshop with Julien.