Published October 13, 2026 Updated October 13, 2026 4 min read

How Accurate Is Throne? We Compared Our AI With Five Gastroenterologists

Scott Hickle
CEO & Co-Founder of Throne

We gave five gastroenterologists the same 278 photos of stool. We asked each of them, working alone, to score every photo on the Bristol Stool Scale, using the standard seven-point reference chart.

They didn't agree with each other as much as you'd expect. On the question of exactly which of the seven Bristol types a given stool was, specialists didn’t consistently land on the same number.

Then we gave our AI the same 278 photos and asked it to do the same.

The study was accepted for presentation at the American College of Gastroenterology's Annual Scientific Meeting, the leading clinical gastroenterology conference in the country. It's our first clinical validation work. Here's what it found, and what it didn't.

How accurate is Throne?

This is one of the questions we get most often, and it has a complication.

To measure accuracy in the strict sense, you need a correct answer to check against. For stool form, there isn't one. No lab test comes back saying "that was a Type 3." The reference standard is expert judgment, and expert judgment varies.

So the question we could actually answer isn't "is the model right." It's this: does the model agree with gastroenterologists about as closely as gastroenterologists agree with each other?

The scale everything runs on

The Bristol Stool Scale is the standard language of gut health. It has seven types, from hard separate lumps at one end to entirely liquid at the other.

Clinicians use it to describe bowel patterns. Clinical trials use it as an outcome measure. The diagnostic criteria for IBS use stool form when sorting the condition into subtypes.

In everyday practice, it runs almost entirely on you remembering.

It looks like this: your doctor asks what your stools have been like. You think back over three weeks, average it in your head, and pick a number off a chart on the wall. That becomes part of the picture your clinician works from.

Physicians already know this is a weak link. A small 2022 study of people with IBS-D found that an AI reading stool photos classified stool form more accurately than the patients' own reports, using expert gastroenterologists as the reference. The obvious fix is to keep a diary, but that has its own well-documented problem. People abandon them, and entries that do get made are often filled in from memory right before the appointment.

So the case for automatic tracking is easy to make. The harder question is whether the automatic version is any good.

What we tested

We pulled 278 de-identified stool images collected from Throne devices during beta testing. The images were balanced across all seven Bristol types based on an initial gastroenterologist label, so no single category could dominate the results.

Five gastroenterologists scored every image independently, using a standard reference chart. Our AI model scored the same set. The model had never seen any of these images during training.

Then we compared everyone against everyone. Gastroenterologist against gastroenterologist. Model against each gastroenterologist. We used the same yardstick for both.

What we found

Researchers measure this kind of agreement with a number called Cohen's kappa. It accounts for the fact that two people will land on the same answer by luck some of the time. A kappa of 1.0 is perfect agreement. A kappa near 0 means you're doing no better than chance.

Averaged across every pair of physicians, agreement on the exact Bristol type was 0.374.

Our model, compared against those same physicians, scored 0.396.

Those two numbers are close, and Throne’s model is marginally higher.

Both numbers rise when you account for how far apart disagreements were, so that calling something a 3 when someone else called it a 4 counts as a smaller miss than calling it a 2 when someone else called it a 6. Scored that way, the physicians averaged 0.594 and Throne’s model reached 0.610.

That pattern held throughout. When the model differed from the physicians, it was usually off by one neighboring category rather than landing somewhere unrelated, which is the same way the physicians differed from each other.

In plain English, the AI classified stool about as consistently with gastroenterologists as gastroenterologists classified it with one another. And when it disagreed, it usually missed by one neighboring Bristol type rather than making a wildly different call.

What comes next

Stool form is the foundation - it's the measurement the rest of gut health gets built on, which is why we validated it first and why we're publishing the result rather than asking anyone to take our word for it.

From here the clinical program runs through hydration and urinary function, with larger prospective studies on stool classification alongside them. Our long-term goal hasn't changed: building the smoke detector for colorectal cancer and cancers of the urinary tract.

What are you looking for?