The Pulse
AI echo analysis identified 78% of confirmed HFpEF cases in study
Adding AI analysis of an echocardiogram clip to a clinical score identified 78% of patients with confirmed HFpEF, versus 48% using the score alone. The study appears in a Journal of Cardiac Failure issue focused on AI in heart failure care.

Adding AI analysis of an echocardiogram clip to a clinical score identified 78% of patients with hemodynamically confirmed heart failure with preserved ejection fraction, compared with 48% using the score alone, according to the Heart Failure Society of America. The study appears in the November 2026 issue of the Journal of Cardiac Failure, which focuses on AI in heart failure.
The condition is known as HFpEF. Researchers applied the EchoGo HF model to resting echocardiograms from 280 outpatients referred for suspected HFpEF. Every patient also underwent exercise hemodynamic testing, which the society describes as the diagnostic gold standard.
Testing a two-step approach
The study found that the AI model added the most value for patients with low or intermediate H₂FPEF clinical scores, where diagnosis is most difficult. The researchers tested a two-step approach: calculate the score first, then use AI when the score is inconclusive.
An AI-detected HFpEF result also independently predicted elevated filling pressures and reduced exercise capacity, the researchers found. The society said that finding suggests the model reflects disease severity as well as diagnosis.
The paper, by Usman and colleagues, addresses how to apply an echocardiography-based AI algorithm in a referral population. The journal is the official publication of the Heart Failure Society of America.
Using language models to examine prescriptions
In another study in the issue, investigators used clinical notes from 599 patients seen in cardiology clinics to optimize OpenAI’s o1 language model. They examined documented reasons why patients with heart failure were not prescribed an SGLT2 inhibitor, a treatment for which many patients qualify under guidelines.
The model classified the documented reason for non-prescription with 84.7% accuracy, according to the society. Physician-attributed factors were the most common theme, appearing in 28.7% of notes.
The leading specific reasons were a decision to defer prescribing to a future visit, at 15.4%, and a physician’s judgment that the patient was low-risk or unlikely to benefit, at 7.4%. The society said the findings point to concrete targets for improving guideline-directed prescribing.
Calls for studies in care
The issue’s guest editors argue that AI needs evaluation beyond retrospective studies and model demonstrations. “Promise alone is insufficient,” they wrote, calling for rigorous study of where AI adds value and how it can be implemented safely and effectively.
They called for prospective studies, including randomized trials where feasible, to test effects on clinical decisions, treatment delivery, outcomes, equity and clinician workload. The guest editors are Jonathan W. Cunningham, Elena Donald, Nir Uriel and Ankeet S. Bhatt.
A review in the issue examines uses ranging from AI-assisted detection in ECGs and imaging to electronic health record risk assessment, decision support, remote medication titration and wearable monitoring. The review authors concluded that the strongest evidence favors AI incorporated into active clinical workflows that guide testing, treatment and follow-up, rather than tools used only for passive alerts.