After Chicago: Did AI at AHA 2026 Pass Its Own Test?

Before the American Heart Association's Scientific Sessions opened in Chicago, I proposed a test for the artificial intelligence studies on the program, in my MarketScale conversation ahead of the meeting, “AI Takes Center Stage at AHA 2026”: do they move beyond showing that an algorithm is accurate to demonstrating that it changes treatment, improves outcomes, or reduces cost? And were they tested in the populations where the need is greatest [1]?

As AHA 2026 concludes its four days in Chicago (November 6 to 9), that test returns a mixed verdict. And perhaps the most instructive results for anyone building AI for heart care were not the AI studies at all.

Three things happened in Chicago that bear directly on the argument my co-authors and I made in our policy brief on precision cardiovascular medicine [2]: that innovation is a public good and health equity a universal right, and that precision tools should strengthen, not replace, the evidence-based practices on which cardiovascular health depends.

The dedicated AI late-breaking session delivered workflow evidence without, as yet, hard clinical outcome evidence. It showed that algorithms can read, triage, and give clinicians back time, which is itself an early signal of benefit, while leaving open whether patients fare better as a result.

In contrast, two of the most anticipated drug trials of the year lowered their targeted biomarkers exactly as intended and changed nothing measurable for patients on their primary clinical endpoints. It was a reminder that engaging a biological target is not the same as delivering a clinical benefit.

And the interventions that actually lowered blood pressure and improved how patients felt and functioned in Colombia, Jamaica, Brazil, and India were team-based primary care, a single combination pill, and a yoga mat.

The lesson from Chicago is not that AI has failed. It is that technological performance is only the beginning of clinical value. The foundation of precision medicine remains the care systems that translate information into action, particularly for the populations that need them most.

The scorecard

Each of the AI studies below advanced the field in a specific way. None yet established that patients experienced better hard clinical outcomes as a result, and that is the gap the next generation of studies must close.

AI ECHO INSIGHT

• What it showed: 1,200 heart ultrasounds at 21 Kaiser Permanente Northern California medical centers; AI-drafted reports needed major revision as often as human drafts (50.6% vs. 49.7%); cardiologist time per study fell from 8.5 to 4.6 minutes [3].
• Did it change treatment, outcomes, or cost? Workflow efficiency, yes. Diagnostic accuracy against a reference standard and patient outcomes were not directly assessed.

WATCH-LV

• What it showed: 2,211 patients wore an Apple Watch and a Galaxy Watch back to back; AI flagged a weak heart pump about equally well on both, but 9% of people got a different answer from the two watches. In home use, acting only when the watch flagged a problem on at least 2 of 3 consecutive days cut false alarms in half (from 18 to 9 of 70 healthy people), at the price of overlooking 2 additional people who did have the condition (31 of 38 detected instead of 33) [4].
• Did it change treatment, outcomes, or cost? Diagnostic screening and a testing rule; no demonstrated improvement in clinical outcomes.

PARAMOUNT

• What it showed: 530 patients with chest pain and high LDL cholesterol, randomized to AI measurement of artery plaque on CT versus a standard stress test; AI-guided patients had the better result twice as often under the trial's composite win-ratio analysis (2.00, 95% CI 1.37–2.56) at 180 days. Cholesterol-lowering medication use rose from 31% to 75% in the AI group, against 32% to 47% with the stress test, and fewer AI-group patients underwent catheterization without an obstructive lesion [5].
• Did it change treatment, outcomes, or cost? Treatment decisions, yes. Hard clinical outcome benefit remains unproven.

PROVAR+

• What it showed: AI reading of ECGs sent through Brazil's public telehealth network to find structural heart disease; in an earlier report it discriminated disease better than the standard rule-based read (AUC 0.73 vs. 0.52) and found half of all cases in 19 days instead of 81 with the same echo capacity [6].
• Did it change treatment, outcomes, or cost? Faster diagnostic triage; downstream treatment and outcomes remain to be established.

PARAMOUNT: turning a picture into treatment

PARAMOUNT came closest to the test. It is a randomized trial of AI-enabled, plaque-based care. In 530 patients with chest pain and elevated LDL cholesterol, it did what imaging in prevention is supposed to do: it turned a picture into treatment.

Instead of a treadmill stress test, patients had a CT scan of the heart arteries, and an algorithm measured the amount and type of plaque in the vessel walls. That gave patients and clinicians a clearer answer about coronary disease, more than doubled the proportion of patients receiving cholesterol-lowering therapy, from 31% to 75%, against an increase from 32% to 47% with the stress test, and sent fewer patients to the catheterization laboratory without a blockage to treat [5].

The outcome measure was a composite that included softer clinical and management measures, the design was open-label, and the company that makes the software ran the trial, so independent scrutiny of the full publication, expected in early 2027, will matter. But the direction is the right one: AI applied to imaging as an engine for prevention rather than simply for procedures.

AI ECHO INSIGHT: the time dividend

AI ECHO INSIGHT is the cleanest workflow design in the group. An echocardiogram, the ultrasound of the heart, is read first by a technologist and then finalized by a cardiologist. In this trial, reviewing cardiologists were blinded to whether a preliminary report came from the AI, a technologist, or a colleague.

The AI drafts were revised no more often than the human ones and took approximately half the time to finalize, 4.6 minutes against 8.5 [3]. Those are minutes a cardiologist could spend with the patient in the room, or reading studies from a clinic that has no cardiologist at all.

INSIGHT did not measure diagnostic accuracy against an independent reference standard or what happened to patients, and the investigators acknowledged that limitation. What it did measure, it measured well.

WATCH-LV: a template for monitoring outside the clinic

WATCH-LV showed that a consumer-device model for detecting a weakened heart pump can travel across manufacturers, Apple and Galaxy, and that repeated readings at home can be made more reliable than a single one [4]. That is a promising template for monitoring outside the clinic.

The population was Korean patients already referred for an echocardiogram, so the next step is a community cohort where the disease is rarer and a false alarm has different consequences.

PROVAR+: the precision-medicine paradox

PROVAR+ is the only study on the list conducted in the kind of setting our brief is about, and it embodies one of its central principles: build on existing public infrastructure and non-physician staff rather than around them.

Brazil's Telehealth Network of Minas Gerais already moves ECGs from primary care clinics to remote readers; adding an AI model turned that flow into a triage system that, in the earlier report, identified half of structural heart disease cases in 19 days instead of 81 with the same echo capacity [6].

It also carries the most sobering number of the session. The model's discrimination in Minas Gerais, an AUC of 0.73 on a scale where 0.5 represents chance and 1.0 perfect discrimination, was well below what it achieved in the US health system where it was built.

That is the precision-medicine paradox in a single statistic: tools may perform less well exactly where they are needed most, unless they are validated, adapted, and recalibrated for those populations.

When the target moves and the patient does not

The two largest disappointments of the meeting had nothing to do with AI and a great deal to teach it. Both trials targeted a blood marker that has long been linked to heart attacks and strokes.

ZEUS went after inflammation, measured by high-sensitivity C-reactive protein (hs-CRP). It randomized more than 6,300 patients with heart disease, chronic kidney disease, and elevated hs-CRP to monthly injections of ziltivekimab or placebo. Inflammation fell as designed. Heart attacks, strokes, and cardiovascular deaths did not: the rate was essentially identical in the two groups, a hazard ratio of 0.99 [8].

Lp(a)HORIZON went after lipoprotein(a), a cholesterol-carrying particle set largely by genetics that raises cardiovascular risk and that diet and statins barely affect. It randomized about 8,300 patients with established heart disease and high Lp(a) to pelacarsen or placebo. Lp(a) fell 78% within a year. Over five years, 503 patients on the drug and 511 on placebo had a major cardiovascular event, about 12% in each group [9].

Two readings are tempting and both are wrong. The first is that inflammation and Lp(a) do not matter. The second is that these trials prove the drugs can never work.

A null result for a therapy added on top of everything else we already do, in patients who already have established disease, does not settle whether the pathway matters. HORIZON patients were on intensive cholesterol-lowering therapy and had carried high Lp(a) for decades before the trial began. Genetic studies suggest that the magnitude and duration of Lp(a) reduction may matter, including whether intervention begins before advanced disease develops. ZEUS enrolled a kidney-disease population with substantial competing cardiovascular and noncardiovascular risks.

The honest reading is that the question stays open, that the next drugs and the next populations will have to answer it, and that nothing in these data licenses optimism without evidence.

What has not changed is what these markers are for. Elevated hs-CRP and elevated lipoprotein(a) still help identify patients at higher cardiovascular risk, and both remain risk-enhancing factors in American cardiology guidelines.

What Chicago changed is how we should act on them. They are signals to be stricter and more intensive about proven treatment and prevention, not proof that lowering those markers with these particular medications reduces heart attacks and strokes.

For a patient with high Lp(a) or persistently elevated hs-CRP, the attention shifts back to the things with demonstrated benefit: a healthy diet, regular physical activity, not smoking, healthy sleep and weight, and tight control of blood pressure, blood sugar, and LDL cholesterol, the eight components the AHA calls Life's Essential 8. A risk marker is a reason to act on the evidence we have, not a reason to wait for the evidence we want.

The lesson for AI

The lesson for AI follows directly. The accuracy statistics that diagnostic algorithms report, the AUC above all, are the software equivalent of a blood marker that fell: they show the model did what it was built to do, separating people with disease from people without, in the population where it was tested.

They do not show that a positive result reached a clinician, that the clinician acted, that the action was right, or that the patient lived longer or better for it. Every AI late-breaker in Chicago reported accuracy, treatment decisions, or a workflow proxy; none established a reduction in hard clinical events attributable to the algorithm.

We would not consider biomarker reduction alone sufficient evidence of clinical benefit from a drug after ZEUS and HORIZON. We should not treat algorithmic accuracy alone as sufficient evidence of patient benefit, least of all for populations where the pathway from result to action is already fragile.

What moved the needle for patients

The clearest winner of the meeting for underserved patients may have been a yoga mat. The interventions that produced meaningful results in the populations our brief is about were low-technology, system-level, and in several cases led from Latin America and India.

Yoga-CaRe HF

Yoga-CaRe HF, from India's Centre for Chronic Disease Control, randomized 3,492 patients with heart failure and a weakened heart pump at 16 centers to a structured, culturally grounded rehabilitation program delivered by certified yoga practitioners, or to enhanced standard care.

Over 15 months, the combined rate of death, hospitalization, or stroke was 21.0% against 22.8%, a difference that could be due to chance. On outcomes patients themselves rank highly, the program performed better: quality of life improved by about 8 points on a standard heart-failure questionnaire, patients walked 25 meters farther in six minutes, and their overall health status improved [13].

A neutral event rate is not a failure. Fifteen months is a relatively short window to change deaths and hospitalizations, and rehabilitation was added to, not substituted for, the medications patients were already taking.

What the trial showed is that cardiac rehabilitation, which remains unavailable to many of the world's heart failure patients, can be delivered by non-physicians, in a culturally appropriate format, and help people feel and function better. That is precision medicine's foundation in action.

CATCH

CATCH, published in JAMA as the meeting opened, randomized 40 primary care clinics in Colombia and Jamaica to team-based care built around a single-pill antihypertensive combination or to standard care.

At 18 months, systolic pressure fell 19.1 mm Hg with the intervention and 9.0 mm Hg with usual care, an adjusted difference of 9.7 mm Hg (95% CI 7.0–12.4). Control to below 130/80 mm Hg was reached by 69.5% of intervention patients and 29.2% of controls [10].

For a lay reader, the scale of that matters: a sustained 10 mm Hg reduction in systolic pressure has been associated with substantial reductions in cardiovascular events across large blood-pressure trials, and the intervention more than doubled the share of patients whose pressure was controlled.

No algorithm was involved. The ingredients were physician and nurse training, health coaching, home monitoring, and audit with feedback: two or three drugs in one tablet so there is one pill to remember, a nurse who follows up, and a home cuff. This is the WHO HEARTS model our brief points to, tested in a randomized trial.

Six months after the program ended, the difference had narrowed to 5.7 mm Hg, which is the implementation lesson in one number. The system carried the effect, and the system has to be sustained.

OPTIMAL-STROKE and OPTIMAL-DIABETES

OPTIMAL-STROKE and OPTIMAL-DIABETES, run in Brazil by Otavio Berwanger's group, together enrolled nearly 14,000 patients and asked how low blood pressure should go in two of the highest-risk groups in cardiology.

After a stroke, pushing the top blood-pressure number down to 120 instead of the usual 140 meant that over four years about 11 in 100 patients had a heart attack, stroke, cardiovascular death, or hospitalization, compared with about 13 or 14 in 100 treated to the standard target: roughly one event prevented for every 40 patients treated more aggressively. The trade-off was real: more patients in the intensive group saw their kidney function decline, which is why the authors stress that this approach needs close monitoring rather than a blanket lower target [11].

In type 2 diabetes, the same target produced a smaller difference, 6.0% against 6.7%, that did not reach statistical significance [12]. Here too, a non-significant difference on top of modern therapy is not evidence of no benefit; the investigators themselves cautioned against that reading. Both trials are practice-informing results generated outside the high-income infrastructure most guidelines rest on, and both show what Latin American trial networks are ready to deliver.

Set these beside the AI session. CATCH moved systolic pressure ten millimeters in two middle-income countries. INSIGHT saved about four minutes of a cardiologist's time per study in Northern California. Both are real, and the second can serve the first: the minutes AI returns are only worth as much as the care systems they are reinvested in.

Task-shifting that worked

A quieter study offered the most direct evidence for the model our brief describes. Within the RURAL cohort, Duke investigators trained 14 nurses, nursing assistants, and CT technologists with no prior ultrasound experience for two hours, then sent them into ten rural counties in Alabama, Kentucky, Louisiana, and Mississippi with AI-guided echocardiography on a mobile unit.

Across 4,597 adults, 98% of studies captured all six standard views, core-lab readers judged more than 99% adequate for measurement, and the median scan took 9.5 minutes. Critical findings, most often a heart pumping below its normal strength, were identified in about 3% and returned to participants within three days in 99% of cases [14].

This is the domestic mirror of the Uganda rheumatic heart disease program our brief cites: non-experts, AI acquisition guidance, central expert review. It also shows how far the acquisition problem has come. A nurse with two hours of training can now produce a research-grade echocardiogram in under ten minutes, which means the scarce resource is no longer the sonographer but the pathway that follows a result.

There was no sonographer comparator, and the study did not follow whether a returned finding changed anyone's care [14]. Those are the next studies to run, and the infrastructure to run them now exists.

The road ahead: turning minutes into patients

INSIGHT poses the most hopeful question of the meeting. If AI cuts a cardiologist's time per echocardiogram from 8.5 to 4.6 minutes, what do we do with the minutes? Multiplied across the millions of heart ultrasounds performed every year, the answer is thousands of clinician-days.

Those can become a longer conversation in the exam room, a same-week read for a rural clinic that waits weeks today, or a cardiologist in São Paulo or Nairobi reading for a district hospital that has never had one. The technology is indifferent to which. The design choices around it are not.

Our brief argues that this is where equity is decided: not in the model but in the implementation framework around it, built with local clinicians, interoperable and privacy-respecting, with training and task-shifting to nurses and community health workers, and evaluated continuously after deployment [2].

Chicago showed each of those elements in isolation. RURAL showed task-shifted acquisition. PROVAR+ showed deployment on public infrastructure, and the recalibration it demands. INSIGHT showed the time dividend. PARAMOUNT showed an algorithm changing treatment. The road ahead is to assemble them into one pathway and test it, end to end, where a result most often fails to become an action.

That is also where the policy levers in our brief become concrete rather than abstract. A pilot fund for precision cardiovascular tools in low- and middle-income settings, regional centers that can validate and recalibrate models on local data, and payment tied to outcomes rather than to reads are what turn a four-minute saving into a patient seen.

The session titled "Are We Ready for AI-Optimized Care?" asked clinicians, health systems, and policymakers that question directly. On the evidence from the late-breaking program, the answer is: ready to measure accuracy and time, and ready now to design the trials that measure benefit.

The device in the patient's pocket

WATCH-LV did the right thing by validating across two manufacturers rather than one premium device, and its home-monitoring results show why what happens after the alert matters as much as the alert itself. Among people whose watch consistently flagged a problem on the days they recorded, 81% had confirmed heart-pump weakness on echocardiography; among those whose readings were mixed, only 40% did [4]. A single buzz on the wrist is a prompt, not a diagnosis.

In a system with echo on demand, a false positive costs an appointment. In a system where PROVAR+ measured the wait to find half of true cases in months, a false positive costs someone else's echo slot.

Cross-manufacturer validation is a real step toward equity of access to the device, and the serial-reading rule is a real step toward making a wrist-worn screen trustworthy. The step after that is to pair the watch with a triage model like PROVAR+ so that a positive reading in a community without a cardiologist knows where to go next.

What to watch next

The test I set before the meeting stands, and the next year will give several chances to meet it.

• LIFE-AI's full results, when published, should be read for whether plaque quantification adds to calcium score and stenosis category, and for subgroup performance across the 269,000 patients its scale permits.
• ARTEMIS after myocardial infarction will show whether IL-6 inhibition has any outcome signal left after ZEUS, with results expected in the first half of 2027. HERMES, the companion trial in heart failure, was stopped for futility in September 2026, as STAT reported on September 7.
• OCEAN(a)-Outcomes and ACCLAIM-Lp(a) now carry the lipoprotein(a) hypothesis alone.
• ANDES, a community health worker trial for hypertension in Puno, Peru, closes enrollment this year.
• The second World Congress on Rheumatic Heart Disease meets in Perth on November 11 to 14, in the year the World Heart Federation has dedicated to neglected cardiac diseases [1].

For AI specifically, the ask is unchanged and now sharper: pragmatic randomized trials with clinical endpoints, run in the populations the tools are meant to serve, with performance reported by sex, age, race and ethnicity, and socioeconomic group before deployment. CATCH shows what that evidence looks like for a pill and a care team. There is no reason an algorithm should be held to less.

Conclusion: connecting tools to care

Precision cardiovascular medicine can become the new standard for everyone, everywhere, but only when it is designed with equity from inception and built to strengthen, not replace, the evidence-based care that Chicago showed still does the heaviest lifting [2].

The acquisition side of that design is well under way: smartwatches that work across brands, mobile echo in the hands of nurses, tele-ECG riding on public networks, and CT that turns plaque into prevention.

The harder and more exciting work is now in reach: connecting those tools to the team-based care, the single pill, and the yoga mat, and proving, in the populations that carry the burden, that the whole pathway makes patients better.

References

1. Reyeszumeta R. AI takes center stage at AHA 2026. Experts warn it must be built for everyone. MarketScale, October 31, 2026. Interview with JR Medina-Inojosa. https://www.marketscale.com/industries/healthcare/ai-aha-2026-built-for-everyone-medina-inojosa
2. Medina-Inojosa JR, Rodriguez F, Lopez-Jimenez F, Sperling LS. Precision cardiovascular medicine: global imperative for equity through innovation. Front Cardiovasc Med. 2026;13:1835607. doi:10.3389/fcvm.2026.1835607 https://www.frontiersin.org/journals/cardiovascular-medicine/articles/10.3389/fcvm.2026.1835607/full
3. American Heart Association. AI-interpreted echocardiograms comparable to human reports, cardiologists' time cut by about 50%. Late-breaking abstract 28202, Scientific Sessions 2026. https://newsroom.heart.org/news/ai-interpreted-echocardiograms-comparable-to-human-reports-cardiologists-time-cut-by-about-50
4. American Heart Association. AI analysis of ECG data measured by smartwatch may help detect reduced heart pumping function. Scientific Sessions 2026. https://newsroom.heart.org/news/ai-analysis-of-ecg-data-measured-by-smartwatch-may-help-detect-reduced-heart-pumping-function
5. Cleerly. PARAMOUNT, the first randomized trial of AI-enabled plaque-based care, meets primary endpoint in late-breaking results at AHA 2026. Press release, November 7, 2026. https://finance.yahoo.com/healthcare/articles/paramount-first-randomized-trial-ai-154500139.html
6. Artificial intelligence-enabled electrocardiography to triage echocardiography for structural heart disease diagnosis in a low-resource setting. Am J Prev Cardiol. 2026;27:101539. https://pmc.ncbi.nlm.nih.gov/articles/PMC13261225/
7. TCTMD. What's going to be hot at AHA 2026. https://www.tctmd.com/news/whats-going-be-hot-aha-2026 Radcliffe Cardiology. AHA Scientific Sessions 2026 late-breaking trials revealed. https://www.radcliffecardiology.com/news/aha-scientific-sessions-2026-late-breaking-trials-revealed
8. BioPharm International. Novo Nordisk's ziltivekimab fails to reduce cardiovascular events in phase 3 ZEUS trial. https://www.biopharminternational.com/view/novo-nordisk-ziltivekimab-zeus-phase-3-cardiovascular
9. American College of Cardiology. Lp(a) HORIZON: pelacarsen vs placebo in patients with high Lp(a). November 7, 2026. https://www.acc.org/Latest-in-Cardiology/Articles/2026/11/06/14/37/sat-10am-lpa-aha-2026
10. Implementation of single-pill combination therapy through team-based care for hypertension control in resource-constrained settings: the CATCH cluster randomized clinical trial. JAMA. Published online November 8, 2026. https://jamanetwork.com/journals/jama/fullarticle/2856399
11. HCPLive. AHA 2026: intensive blood pressure control reduces events after ischemic stroke (OPTIMAL-STROKE). https://www.hcplive.com/view/aha-2026-intensive-blood-pressure-control-reduces-events-after-ischemic-stroke
12. American College of Cardiology. Intensive BP lowering, yoga-based cardiac rehab highlighted among late-breaking science at AHA 2026 (OPTIMAL-DIABETES, Yoga-CaRe HF). https://www.acc.org/latest-in-cardiology/articles/2026/11/06/14/37/sat-5pm-yoga-aha-2026
13. American Heart Association. Yoga-based cardiac rehab improved quality of life in people living with heart failure. Late-breaking abstract 21982. https://newsroom.heart.org/news/yoga-based-cardiac-rehab-improved-quality-of-life-in-people-living-with-heart-failure
14. American Heart Association. Non-experts with AI guidance and off-site review perform quality mobile rural heart scans. Abstract MDP2290. https://newsroom.heart.org/news/non-experts-with-ai-guidance-and-off-site-review-perform-quality-mobile-rural-heart-scans

Late-breaking abstracts are preliminary until published in peer-reviewed form. Figures above are as reported by the sources cited on November 9, 2026.