A randomized test, not another model demo
PARAMOUNT has given clinical AI a more consequential kind of validation: not just whether an algorithm can identify coronary plaque, but whether physicians using AI-derived plaque information manage patients differently and, by the trial’s chosen measure, more effectively. Cleerly said on November 7 that the study met its primary endpoint in late-breaking results presented at the American Heart Association Scientific Sessions 2026 in Chicago .
That matters because much of medical AI’s evidence base still lives in the world of retrospective accuracy studies, reader-assistance experiments and workflow pilots. Those studies can be useful, but they often stop before the harder question: does the tool improve care when inserted into an actual clinical pathway? PARAMOUNT was designed to ask that question in patients with stable chest pain and suspected coronary artery disease, comparing AI-guided plaque-based care with usual care guided by functional stress testing .
The result is therefore important even with major caveats. The trial does not yet show that AI-guided care prevents myocardial infarctions or deaths. Its primary endpoint was a hierarchical 180-day composite built around diagnostic certainty, LDL-C reduction and the efficiency of invasive coronary angiography or revascularization referral . Still, for hospitals, insurers and regulators, a randomized signal that an AI tool changes diagnosis and prevention decisions is a higher rung of evidence than a standalone performance benchmark.
What PARAMOUNT tested
PARAMOUNT enrolled symptomatic outpatients with suspected coronary artery disease, elevated LDL-C of at least 100 mg/dL and no prior history of coronary artery disease . Between January 2025 and February 2026, investigators randomized 530 participants at 21 U.S. sites either to plaque-based care guided by AI-based quantitative coronary CT analysis, including quantitative plaque and plaque-based ischemia assessment, or to usual care guided by functional stress testing .
The late-breaking analysis was based on 523 participants with locked data at the time of analysis . The population is worth noting: women made up 57% of participants, or 300 of the 523 analyzed patients, an unusually strong representation in cardiovascular research, while the median age was 57 and median baseline LDL-C was 127 mg/dL .
In practical terms, the AI-guided arm used coronary CT angiography data processed into a more detailed picture of atherosclerotic plaque: its amount, morphology and composition. HCPLive summarized the comparison as coronary plaque-based care versus usual care, with the AI-enabled arm receiving quantitative assessment of plaque morphology and composition, while usual-care patients were managed after stress-test findings .
The headline numbers
On the prespecified hierarchical composite endpoint at 180 days, plaque-based care was favored over usual care with a win ratio of 2.00, with a 95% confidence interval of 1.37 to 2.56 and p<0.0001 . In unmatched patient comparisons, the plaque-based-care patient had the better result 46.7% of the time, versus 23.4% for usual care .
The biggest contributor was not an event reduction. It was improved certainty about whether the patient’s chest pain was caused by coronary artery disease, with a win difference of 16.6 percentage points . That may sound less dramatic than a mortality endpoint, but diagnostic certainty is not trivial in chest-pain care. It influences whether clinicians escalate testing, send patients to the cath lab, prescribe preventive treatment or reassure patients.
The trial also reported a large difference in preventive therapy uptake. In the AI-QCT-guided arm, lipid-lowering therapy rose from 31% at baseline to 75% after testing, compared with 32% to 47% in the usual-care arm . LDL-C reductions also favored AI-QCT-guided care, according to Cleerly’s release .
A second operational signal came from invasive angiography. Among patients who underwent invasive coronary angiography, 38% in the AI-QCT-guided arm did not go on to revascularization with PCI or CABG, compared with 60% in the usual-care arm . That does not mean the AI arm sent fewer people overall to invasive procedures; Cleerly specifically said plaque-based care did not increase overall rates of catheterizations or revascularizations . The more precise interpretation is that, among those referred, the AI-guided pathway appeared to produce a more targeted cath-lab population.
Why this result is bigger than the software
PARAMOUNT is being read as an AI story, but it is also a cardiology strategy story. Traditional chest-pain evaluation often asks whether there is flow limitation or ischemia. Plaque-based care shifts attention toward the underlying atherosclerotic disease process, including nonobstructive plaque that may still carry risk and may justify aggressive prevention.
That is why the lipid-lowering therapy result is central. If AI-QCT mainly makes invisible or underappreciated plaque more concrete for clinicians and patients, its first measurable clinical effect may be better prevention rather than fewer near-term procedures. Todd C. Villines, PARAMOUNT co-principal investigator, told HCPLive that stress testing can still leave patients and clinicians uncertain about whether plaque is present, and that definitive answers matter when providers order tests .
For insurers and health systems, this is exactly the type of evidence gap that has slowed broad adoption of algorithmic decision support. A model can be FDA-cleared, impressive in validation datasets and popular with clinicians, yet still face skepticism over reimbursement if it cannot show that using it changes management in a useful way. PARAMOUNT does not settle the reimbursement debate, but it supplies a randomized piece of the puzzle: AI-guided plaque assessment was associated with more diagnostic certainty, more lipid-lowering therapy and a favorable 180-day composite .
The caveats are essential
The limitations are not footnotes. PARAMOUNT was prospective, multicenter and randomized, but it was also open-label . That is understandable for a care-strategy trial, because physicians know whether they are using AI-QCT or usual stress-test results. But open-label design can influence behavior, especially when endpoints include diagnostic certainty and treatment decisions.
The findings are also preliminary and have not yet been peer reviewed . Cleerly said full results are expected to be published in early 2027 . Until then, outside clinicians and payers will want to see the full endpoint hierarchy, the distribution of wins, subgroup results, imaging downstream use, medication intensity, adherence and safety signals.
The endpoint itself also deserves careful reading. A win ratio of 2.00 is strong for the prespecified composite, but that composite combines several domains. It should not be translated into “AI cut heart attacks in half,” because PARAMOUNT has not reported such an outcomes result. Superpower Daily’s summary made the same caution explicit: the main measure combines diagnostic, cholesterol-control and invasive-procedure elements and should not be read as evidence that the strategy prevented heart attacks .
A bridge to outcomes trials
The next question is whether this kind of plaque-guided care can improve hard cardiovascular outcomes over longer follow-up. Cleerly is already positioning PARAMOUNT beside a broader program. In its November 5 AHA preview, the company described TRANSFORM as a randomized, event-driven outcomes trial testing an AI-guided coronary artery disease staging system in asymptomatic patients at increased risk for heart attacks, with primary results anticipated in 2029 .
That distinction is important. PARAMOUNT is a clinical strategy validation trial in symptomatic stable chest-pain patients. TRANSFORM is designed to ask whether AI-guided staging and treatment can reduce heart attacks and other cardiovascular events in a different, higher-risk prevention setting . If PARAMOUNT shows that AI can change diagnosis and treatment decisions, TRANSFORM is aimed at the harder question of whether those decisions translate into fewer events.
For now, PARAMOUNT is best understood as a milestone rather than a finish line. It gives clinical AI something more durable than a demo: a randomized comparison under real care conditions, a statistically significant primary endpoint and a plausible mechanism through more targeted diagnosis and prevention. The cardiology skill tree has not reached the final boss of outcomes proof, but it has unlocked a serious new level of clinical validation.
Comments
Be the first to comment.