AI Stool Test Detects 90% of Colorectal Cancers: What the Study Means

An AI-driven stool test built at the University of Geneva detects roughly 90% of colorectal cancers by reading the gut microbiome at the subspecies level, approaching the ~94% sensitivity of colonoscopy and beating today’s at-home stool tests. The proof-of-concept was published in Cell Host & Microbe in 2025 and has been gaining attention again in 2026 as the American Cancer Society updates its screening guidance to lean harder on non-invasive at-home options. The catch: this is a research model, not an FDA-cleared product, and clinical trials with Geneva University Hospitals are only now starting.

Key takeaways

• A 2025 Cell Host & Microbe study used machine learning on stool microbiome data and detected colorectal cancer with ~90% sensitivity, close to the ~94% rate of colonoscopy.
• The breakthrough was resolving gut bacteria at the subspecies level — finer than the species-level catalogs used by older microbiome tests — which surfaced cancer-linked microbial patterns that earlier methods missed.
• Today’s FDA-cleared at-home options sit lower on sensitivity: FIT detects ~74% of colorectal cancers and Cologuard ~92%, per the pivotal trial reported on the Cologuard label.
• This is not yet a product you can buy. A first clinical trial at Geneva University Hospitals is being set up to validate which cancer stages and pre-cancerous lesions the model can actually detect.
• The American Cancer Society’s 2026 guideline update reaffirms that average-risk adults should start screening at age 45 and adds new at-home stool and blood-based options as valid choices alongside colonoscopy.

What the study actually found

The research, led by the lab of Professor Mirko Trajkovski in Geneva’s Faculty of Medicine, did two things in sequence.

First, the team built a more granular map of the human gut microbiome. Older microbiome studies usually stop at the species level — they tell you Escherichia coli is present, for example, but not which strain. Using a sketching-based method the authors call panhashome, the Geneva group catalogued bacteria at an “operational subspecies unit” (OSU) resolution across thousands of public stool samples, capturing intraspecies variation that earlier work flattened out.

Second, they trained a machine-learning classifier on stool microbiome data from people with and without colorectal cancer. According to the press summary from the University of Geneva, the subspecies-level model correctly identified cancer in roughly 90% of cases — much higher than equivalent models built at the species level, and only a few percentage points below the ~94% sensitivity associated with colonoscopy.

The biological reason this matters: in several cases, one subspecies of a gut bacterium was strongly associated with cancer while a sibling subspecies of the same organism was not. Species-level analyses average those out and lose the signal. Subspecies analysis keeps it.

How it compares to existing stool tests

There are already two FDA-cleared at-home stool tests for colorectal cancer screening in the US. Both are good. Neither is as sensitive as the Geneva AI model in early reports.

Test	What it looks for	Reported cancer sensitivity	Reported specificity
FIT (fecal immunochemical test)	Blood proteins (hemoglobin) in stool	~74%	~95%
Cologuard (mt-sDNA)	Blood proteins + tumor DNA markers	~92%	~87%
Colonoscopy	Direct visualization + polyp removal	~94% (the gold standard)	—
Geneva AI microbiome model (research)	Subspecies-level bacterial signature	~90% (proof-of-concept)	Not yet established in trials

Numbers for FIT and Cologuard come from the Cologuard pivotal trial summary, which used colonoscopy as the gold standard across ~10,000 patients. The Geneva number is from a research model that has not yet been through a prospective clinical trial, so it should be read as a promising signal, not a final performance figure.

One important nuance: sensitivity for cancer is not the same as sensitivity for pre-cancerous polyps. Colonoscopy lets the doctor remove polyps before they turn malignant, which is what makes it preventive, not just diagnostic. Stool tests catch cancer that’s already there. The Geneva paper’s clinical trial is specifically designed to map which polyp stages the microbiome model can pick up — that question is still open.

Why this matters now

Two things make 2026 the right moment for this story.

Colorectal cancer is rising in young adults. The American Cancer Society’s 2026 statistics report projects 158,850 new diagnoses and 55,230 deaths in the US this year. Incidence is rising about 3% per year in people aged 20–49 and is now the leading cause of cancer death in adults under 50. One in five new diagnoses is in someone under 55. The full picture of who needs screening is shifting younger, fast.

Most people skip the screening they have. Screening prevalence in adults under 50 is only about 37%, and roughly 3 in 4 colorectal cancers in this age group are diagnosed at an advanced stage. Cheap, accurate, non-invasive screening is the most realistic lever for catching cancers earlier in people who would never book a colonoscopy. The ACS 2026 guideline update explicitly leans into this: it adds at-home stool and blood-based tests as valid choices, while keeping visual exams (mainly colonoscopy) and stool-based tests as the preferred preventive options.

If the Geneva model holds up in trials, it would slot into exactly this gap: a mail-in stool test that’s nearly as sensitive as colonoscopy for cancer, costs a fraction as much, and doesn’t require a prep day.

What this means for tracking your own stool

A microbiome-based screening test analyses what’s in your stool. A photo-based stool tracker like PoopCheck analyses what your stool looks like — Bristol type, color, and consistency over time. These are complementary, not competing.

The signals that show up in a daily stool log are different from the signals a microbiome assay reads, but both matter for catching change early:

• Persistent changes in Bristol type (formed stool becoming chronically loose or pellet-like) are one of the symptom patterns that the ACS lists as a reason to talk to a doctor, particularly when paired with bleeding, weight loss, or anemia.
• A new pattern of narrow or pencil-thin stool that lasts more than a couple of weeks deserves attention — see our deeper look at pencil-thin stools and when to worry.
• Black, tarry, or persistently bloody stool is always a red flag — covered in detail in black stool causes and in our overview of colon cancer symptoms.

None of this replaces a real screening test. It’s the slower-moving, daily-resolution layer that helps you and your doctor notice when something has actually changed.

What this doesn’t mean

A 90% sensitivity headline is genuinely exciting, but the gap between a research model and a clinical product is large. A few caveats worth keeping in mind:

• No regulatory clearance yet. The Geneva model is a published research method, not an FDA-cleared (US) or CE-marked (EU) diagnostic. It is not available at your doctor’s office.
• The 90% figure may shift. Sensitivity reported in a model-training and validation paper is not the same as sensitivity in a prospective, multi-site clinical trial. The number could rise or fall once the Geneva University Hospitals trial reports.
• Polyp detection is the harder problem. Even the best stool tests detect fewer pre-cancerous lesions than they do established cancers. Cologuard’s sensitivity for advanced adenomas is around 42% according to its pivotal trial — much lower than its cancer sensitivity. The same gap likely applies to microbiome models.
• Microbiome tests for non-cancer “gut health” are a different category. A test built and trained for colorectal cancer detection is not the same product as a direct-to-consumer “what does your microbiome say about you” kit. We’ve covered why most of those tests are not worth the money in our review of gut microbiome testing in 2026.

In other words, this is a genuine advance in research, not an excuse to skip your screening test today.

FAQ

Can I get the Geneva AI stool test now? No. It is a research model published in Cell Host & Microbe, not a commercial product. A first clinical validation trial is being organized in collaboration with Geneva University Hospitals.

Is this better than a colonoscopy? For detecting established cancer, the model reaches roughly 90% sensitivity versus the ~94% of colonoscopy. But colonoscopy is also preventive — the doctor can remove pre-cancerous polyps during the exam, which a stool test cannot do. The two tools answer different questions.

How does it compare to Cologuard and FIT? The headline cancer sensitivity (~90%) is higher than FIT (~74%) and slightly lower than Cologuard’s reported ~92%. None of these tests are head-to-head trialed against the Geneva model yet, so direct comparisons should be cautious.

Should I still get screened at 45? Yes. The American Cancer Society’s 2026 guideline update reaffirms screening from age 45 to 75 for average-risk adults. Talk to your doctor about which method fits you — colonoscopy, FIT, mt-sDNA (Cologuard), or one of the newer at-home or blood-based options.

Does a regular at-home microbiome test screen for cancer? No. Direct-to-consumer microbiome tests like ZOE, Viome, or Ombre are not cancer screening tests. They generate microbial profiles, not validated cancer predictions, and a 2024 international consensus concluded current consumer tests lack established clinical validity for medical use.

What stool changes should make me see a doctor? Persistent change in bowel habits, narrow or pencil-thin stool that lasts more than a couple of weeks, visible blood, dark or tarry stool, unexplained weight loss, fatigue, or iron-deficiency anemia. See our full guide to colon cancer symptoms and when to see a doctor.

The bottom line

A subspecies-level microbiome model that detects 90% of colorectal cancers from a stool sample is a real advance — and one of the most credible signals yet that AI-driven stool analysis is going to play a serious role in cancer screening. It is not, today, a replacement for getting screened at 45 the way your doctor recommends. The most useful thing most adults can do in 2026 is the boring thing: pick a screening method you’ll actually complete, log what your stool looks like between visits, and pay attention to changes that persist.

If you want a low-friction way to capture that day-to-day signal, the PoopCheck app analyses Bristol type, color, and consistency from a photo and shows you the trend over time — so when something does change, you have data instead of memory.

Sources

1. Tričković M, et al. Subspecies of the human gut microbiota carry implicit information for in-depth microbiome research. Cell Host & Microbe, 2025. (Lead investigator: Mirko Trajkovski, University of Geneva.) PubMed entry.
2. University of Geneva. Gut bacteria can reveal colorectal cancer. UNIGE Faculty of Medicine press release.
3. American Cancer Society. American Cancer Society Updates Colorectal Cancer Screening Guideline. 2026.
4. Wolf AMD, et al. Colorectal cancer screening: an update to the American Cancer Society guideline, 2026. CA: A Cancer Journal for Clinicians, 2026.
5. Siegel RL, et al. Colorectal cancer statistics, 2026. CA: A Cancer Journal for Clinicians, 2026.
6. Exact Sciences. Cologuard accuracy, sensitivity, and specificity (pivotal trial data).