Clinical Before Technical: Building Real-world Endpoints for Hepatology

A Familiar Problem

Oncology ran into this problem first. RECIST depends on scheduled imaging and often on central review. A typical real-world patient chart contains neither. So the field defined a related but different concept, real-world progression, and tested it against trial results and survival. One of many hard parts was deciding what progression looks like in a chart.

For hepatology, the parallels run deep. The FDA recently accepted a letter of intent to evaluate liver stiffness measured by VCTE as a reasonably likely surrogate endpoint in non-cirrhotic MASH with moderate-to-advanced fibrosis. What the field needs next is real-world longitudinal evidence showing that improvements in liver stiffness actually predict a reduction in complex liver-related events like decompensation. Answering that question requires a dataset built to capture both the liver stiffness trajectory and those critical events with depth and validity.

Where The Usual Sources Fall Short

Structured EHR fields and claims are not enough. A code for ascites or encephalopathy usually has no onset date and no grade, and may not even get entered at the time of initial finding. A claim for a GI bleed admission doesn't reliably tell you whether the bleed was variceal. The detail is in the unstructured data including the notes and the endoscopy, imaging, and elastography reports.

Having access to the unstructured data isn’t enough to make it usable.

For example, notes aren’t written with research use cases in mind, and a lot of what's in them may not perfectly support real-world data creation. "History of ascites" has limitations. "No evidence of encephalopathy" has the word encephalopathy in it. Findings related to a single event can be described across a progress note, a discharge summary, and a procedure report, each adding a bit more detail to the overall picture.

You can run a model over all of the raw information and get a populated column back, but that doesn't mean you’ve created a good variable or have any idea how it performs. For that, a clinician (ideally a hepatologist) has to decide up front what counts as the event. Then the output has to be checked against charts read by people who know the disease area, and the quality of this process needs to be documented. Skip any step and it becomes hard to defend the variable when there’s a question about how it was made.

What Our Hepatologists Defined

Building a good variable is clinical before it’s technical.

So before we assembled any data, our hepatologists wrote definitions for the concepts hepatology studies depend on, including what a record has to show.

• Decompensation. Ascites, variceal hemorrhage, and overt encephalopathy are defined as separate events, each with its own criteria and its own date. None of them is read off a diagnosis code alone.

• Liver stiffness. Elastography results are kept as dated values rather than a single number per patient. One reading shows where a patient stood on a given day. A series shows how they're changing, which is what a progression study needs.

• Fibrosis and severity scores. The standard serum-based indices and severity scores are built from underlying lab values, so a team can see how each one was calculated.

• Death, transplant, and cardiovascular events. Death competes with nearly every other outcome, and a health system often doesn't see deaths that happen elsewhere, so we don't rely on its record alone. Transplant is a dated event. Heart attack, stroke, and coronary revascularization are also there with the liver outcomes, since cardiovascular disease is the leading cause of death in MASLD.

• Cirrhosis. We confirm it from the clinical record instead of inferring it from a code, and each confirmed case can be traced back to its source.

These are critical building blocks of good hepatology evidence. Endpoints such as time to first decompensation, or stiffness response after treatment start, are constructed from them. They will be related to trial endpoints but not identical, just as real-world progression in oncology RWE is related to RECIST progression. Of course, every one of these variables is limited by what the chart contains so not every data point is available for every patient.

Definitions Are Just The Start

Real-world progression gained credibility through validation. It was benchmarked against trial results, tested for sensitivity to how it was defined and when patients were assessed, and reported openly. Hepatology will need the same: showing that real-world stiffness changes resemble what trials observe, that results hold under reasonable alternative definitions, and that the information can be confirmed at the source.

We are taking on all this work at Pedestal Health.

Until then, these are the types of questions I'd put to anyone developing evidence in hepatology, including us. Which clinical concepts are in the data model as defined variables, and who wrote the definitions? Where did each value come from? How was it confirmed, and will that record be available to the team that has to defend the study?

If the answers come readily, the data may have been built well and with your research question in mind.

If not, you should probably be careful.