Biograph Says 42.9% of Healthy Adults Have Coronary Inflammation on AI CT Analysis. What Holds Up?

A private preventive-health clinic says AI applied to a heart CT scan found elevated coronary inflammation in 42.9% of apparently healthy, low-risk adults, and that none of the usual blood tests or risk calculators caught it. That is a striking headline. It is also a press release about a self-selected clinic population, not a peer-reviewed trial, and the authors themselves say the clinical benefit is unproven. This post explains the claim, what the underlying technology is, and what you can and cannot conclude.

TL;DR

What Biograph reports

According to the Biograph press release, the study was available to clinically eligible members of its Biograph Black tier and combined high-resolution coronary CT angiography with cardiometabolic profiling and Caristo's AI. The reported results:

• 42.9% of participants had elevated coronary inflammation.
• 17.9% had very high inflammation in at least one major coronary artery.
• About 47% of participants with no detectable plaque had elevated coronary perivascular inflammation.
• Routine markers including ApoB, non-HDL cholesterol, HbA1c and hs-CRP, and the AHA PREVENT-ASCVD calculator, did not correlate with the inflammation.

Michael Doney, Biograph's executive medical director, said that "no routine blood biomarker, lipid metric, or clinical risk score could identify these individuals with inflamed coronary arteries," and that the finding may point to "an upstream phase of disease and a window for proactive intervention." He added: "More research is needed to validate this finding and establish the prognostic value and benefit of care directed at coronary perivascular inflammation."

The release itself gives no participant count, ages or threshold, but it links a medRxiv preprint (not peer reviewed) that does. It reports 84 consecutive asymptomatic adults with no prior coronary disease (mean age 53.7, 63.1% men) at two clinical sites. "High" and "very high" perivascular inflammation were pre-specified as any vessel at or above the 75th and 90th percentiles of the FAI-Score, so 42.9% and 17.9% correspond to roughly 36 and 15 people. All four authors are Biograph employees and Biograph sponsored the study; they report no relationships with Caristo. A percentile cutoff also means the "elevated" label is relative by design, not an absolute clinical threshold.

How the technology works

CaRi-Heart is Caristo's AI platform. According to trade coverage of the FDA decision, such as HIT Consultant's July 29 report, the algorithm analyzes the fat tissue around the coronary arteries on a routine coronary CT angiogram to produce a fat attenuation index, a measure intended to reflect inflammation in the vessel wall, and feeds it into a risk score that also estimates 10-year cardiovascular mortality risk. The idea: inflamed arteries change the surrounding fat's appearance on CT, even when the artery itself shows no plaque.

The same coverage says the FDA granted De Novo authorization, describing CaRi-Heart as the first U.S. technology to quantify coronary inflammation from routine CCTA, with support from validation data published in The Lancet and outcomes data from the ORFAN registry. We did not retrieve the FDA database entry or the paper, so treat those details as reported by the company and trade press. CaRi-Plaque, which does quantitative plaque analysis, was cleared earlier according to a March 2025 report.

What is confirmed and what is not

Why 42.9% is not "42.9% of people"

Three reasons to avoid generalizing.

1. Selection. The participants are members of a premium preventive-health tier who chose or were eligible for advanced imaging. They are not a random sample of adults, and they differ in age, income and health behavior.
2. A threshold is a choice. "Elevated" depends on a cutoff. The preprint uses the 75th percentile in any of three vessels, and with only 84 people the figure is imprecise.
3. Prevalence is not outcomes. Finding a marker is different from showing that people with the marker have heart attacks at a higher rate, or that acting on the marker helps. The outcome evidence the company cites comes from other, larger studies (the ORFAN registry and published validation), not from this clinic cohort.

There is also a commercial context. Biograph sells preventive diagnostics, and the release sits on its newsroom. Neither fact makes the result wrong, but both are reasons for caution, and for independent replication.

What an FDA De Novo authorization does and does not mean

A De Novo authorization is the FDA pathway for novel, low-to-moderate-risk devices with no existing comparable product. It means the agency judged the device safe and effective for its stated intended use, with controls. It does not mean every claim made about the device, or every way a clinic uses its output, has been proven. The intended use matters: according to the trade coverage, CaRi-Heart is authorized to quantify coronary inflammation from routine CCTA and to produce a risk estimate. A press release about prevalence in a membership clinic is a different kind of claim from the device's cleared use, and the FDA has not endorsed it.

That is a general lesson for any AI product marketed alongside a regulatory milestone. Ask which sentence the regulator actually reviewed.

How to read a study announced in a press release

Whether the topic is cardiology or a language model, a few questions sort quickly.

1. How many people, and who? The release gives no sample size; the preprint says 84. That is a few dozen people, not thousands.
2. What was measured against what? A claim that blood tests "showed no correlation" needs a correlation statistic and a confidence interval. Absence of a significant correlation in a small sample is weak evidence of no relationship.
3. What is the comparator? Prevalence in a general population, or in people with known disease, would put 42.9% in context. Without a comparator, the figure is a number without a baseline.
4. Who funded and ran it? A clinic that sells the scan, a vendor that makes the software and a press office all have an interest in the headline.
5. What are the outcomes? The study measured a marker, not events such as heart attacks. Marker studies are a first step.
6. Is it reproducible? Look for a preprint, a registry entry or a methods document.

Applying that list here: the sample is small, comparators are missing, outcomes are absent, and the sponsor sells the service. That does not make the finding false. It makes it a hypothesis worth testing, which is also how the medical director framed it.

Why it matters for AI in medicine

This is a good example of a pattern in clinical AI: software extracts a new signal from an imaging test that already exists. The appeal is that the CT scan was acquired anyway, so the AI adds information without extra radiation. The risk is overdiagnosis: if a large share of low-risk people get flagged, you need proof that flagging leads to better outcomes before it becomes a routine screen. We have seen the same questions in other clinical AI coverage:

• Microsoft's Diagnostic Orchestrator scores high on past cases, with real-world validation still ahead.
• Google AMIE in a real clinic shows what a prospective study looks like.
• Mass General Brigham's FAHR model is another AI marker with outcome data behind it.
• Holon S1 and Sophont extract gene-expression signals from pathology slides.

What to do if this affects you

• Do not seek out a scan because of this press release. Ask your clinician whether coronary CT is appropriate for your risk, symptoms and family history.
• Ask what the result would change. If a flagged score would not alter treatment, its value is limited.
• Ask about radiation, contrast and cost. CCTA involves a CT scan and typically contrast dye.
• Look for the paper. If peer-reviewed results appear, check sample size, thresholds, follow-up and conflicts of interest.
• Keep the basics. Blood pressure, lipids, glucose, smoking and exercise still drive risk, even if this study says they did not track the inflammation score.

What to watch next

• Peer-reviewed publication and a larger sample.
• Follow-up outcomes for the flagged participants.
• Whether insurers and guidelines adopt coronary inflammation scoring for low-risk people.
• Independent clinics reporting prevalence in general populations.

Related reading

• Microsoft Diagnostic Orchestrator in Science
• Google AMIE in a real clinic
• Mass General Brigham FAHR face model
• Omnii-1: first AI-designed phage given to a patient
• Nolla Health: AI-issued acne prescriptions in Utah
• Holon S1 pathology gene expression

This is not medical advice. Figures are from a company press release dated October 22, 2026.