A patient waits 8 years for a celiac diagnosis. That is the average delay, not the worst case. The disease affects roughly 1% of the global population, and the diagnostic bottleneck is real: subjective pathologist reads of duodenal biopsies, inconsistent Marsh grading criteria, and a global pathologist shortage that runs 20-30% vacancy in some regions. So when AI advocates argue that machine learning could standardize image analysis and close that gap, I understand the appeal. I just need them to show me the study.
They cannot. Not yet. As of April 2026, there is no peer-reviewed, adequately powered trial validating an AI system as a primary diagnostic tool for celiac disease. Roche's April 2026 investor materials describe AI algorithms for digital pathology workflows, but offer no celiac-specific data, no sensitivity or specificity numbers, no comparison against expert pathologist reads. That is not a product. That is a roadmap slide.
What the Evidence Actually Supports
The honest case for AI in celiac pathology is narrower and more useful than the replacement narrative. Villous atrophy grading, the core task in celiac biopsy interpretation, is exactly the kind of pattern-recognition problem where convolutional neural networks have shown promise in adjacent fields. A 2019 study in Gut demonstrated that AI could classify colorectal polyp histology with accuracy comparable to expert pathologists. That is encouraging. It is also not celiac, not a randomized trial, and not replicated at scale in clinical settings.
The gap between "AI can do this in a controlled research environment" and "AI should replace a trained pathologist in a clinical setting" is where people get hurt. Celiac diagnosis carries real consequences: a lifelong gluten-free diet, downstream monitoring for lymphoma risk, and the psychological weight of a chronic disease label. A false positive from an unvalidated algorithm is not a minor inconvenience.
Proponents of faster AI adoption will correctly point out that delayed diagnosis also causes harm, including years of malabsorption, anemia, and bone density loss. That tension is real, and I am not dismissing it. But the answer to a slow, imperfect human system is not an unvalidated automated one. The answer is a validated automated one, which requires running the trials first.
The Microbiome Distraction
A study published April 21, 2026 on oral microbiome aging scores generated excitement about noninvasive AI-driven health prediction. The Oral Microbiome Aging Acceleration score linked each unit increase to a 5% higher all-cause mortality risk. Interesting work. Completely irrelevant to celiac diagnosis. I mention it because the wellness press has a habit of laundering adjacent AI health findings into broader claims about AI replacing specialists. A machine learning model predicting biological age from an oral rinse tells us nothing about whether AI can reliably grade villous atrophy on a duodenal biopsy slide.
What should actually happen: gastroenterology societies and pathology boards should fund prospective trials testing AI-assisted Marsh grading against expert consensus reads, with pre-registered endpoints and independent validation cohorts. Roche and similar companies developing digital pathology tools should publish their celiac-specific performance data, not bury it in investor presentations. Regulators should require clinical validation before any AI system is positioned as a primary diagnostic tool in this space.
AI will almost certainly improve celiac diagnosis. The technology is plausible, the need is documented, and the image analysis task is tractable. But "almost certainly will" and "has been shown to" are different sentences, and in medicine, only one of them justifies changing clinical practice. The patient waiting 8 years deserves a faster answer. They do not deserve a faster wrong one.
