Every year, hundreds of thousands of people in the United Kingdom undergo a cardiac CT scan, usually to check for blocked arteries or calcium buildup. The scan is read, a report is filed, and the patient moves on. But according to a study published in the Journal of the American College of Cardiology in early 2026 (note: this DOI has not been independently verified by this publication), those same images may hold a hidden forecast: whether the patient will develop heart failure within the next five years.
A team at the University of Oxford, led by Prof. Charalambos Antoniades of the Radcliffe Department of Medicine, has developed an AI model that analyzes the thin layer of fat surrounding the heart on standard CT scans. By extracting dozens of texture and density features from that fat, features invisible to the human eye, the tool identified patients headed toward heart failure with 86 per cent accuracy in a large, independent test group. The study does not specify whether that 86 per cent figure refers to a C-statistic, sensitivity, balanced accuracy, or another metric, so the number should be interpreted with that ambiguity in mind. If the results survive prospective clinical trials, the technology could transform a scan that millions already receive into an early warning system for one of the leading causes of hospitalization worldwide.
What the study found
The AI applies a technique called radiomic phenotyping to epicardial adipose tissue, the fat pad that wraps directly around the heart muscle. Rather than measuring how much fat is present, the model maps subtle patterns in its texture and density. Those patterns, the researchers argue, reflect low-grade inflammation and structural remodeling that precede clinical heart failure by years, long before a patient notices shortness of breath or ankle swelling.
To train the model, the Oxford team used anonymized cardiac CT images from more than 59,000 people. They then tested it on a separate cohort of 13,424 patients drawn from nine different NHS Trusts, with outcome data stretching back roughly a decade. That multi-site validation is significant. Single-hospital AI studies frequently perform well on home data but collapse when applied elsewhere, a problem known as overfitting. Demonstrating consistent accuracy across nine hospital systems, each with its own scanner hardware, patient demographics, and imaging protocols, offers stronger evidence that the 86 per cent figure is not an artifact of one dataset.
According to the university’s announcement, this represents the first time heart failure has been predicted from routine cardiac CT using AI. The claim is notable because heart failure, particularly the subtype known as heart failure with preserved ejection fraction (HFpEF), is notoriously difficult to catch early with conventional tools. Standard screening often relies on echocardiography or blood tests for natriuretic peptides, both of which typically flag the condition only after the heart’s function has already measurably declined.
Why fat around the heart matters
The idea that cardiac fat carries diagnostic signals is not new to the Oxford group. In 2019, the same team reported that a “fat radiomic profile” derived from CT scans could predict fatal heart attacks years in advance, research described as funded by the British Heart Foundation. No direct citation for that 2019 work is available in the current source material, so readers should treat the reference as based on the Oxford group’s own descriptions rather than an independently verified publication link.
Independent research has reinforced the concept. A review co-authored by Prof. Antoniades in the Journal of Cardiovascular Computed Tomography described how perivascular and epicardial fat functions as an in vivo marker of vascular inflammation, with detectable shifts in CT attenuation and texture. Separately, clinical work on perivascular fat attenuation has linked CT-derived metrics specifically with HFpEF, the subtype that standard diagnostics most often miss. These parallel findings from multiple groups suggest that fat-based imaging signatures carry genuine biological meaning, not just statistical noise.
A related but distinct line of research has explored radiomic analysis of coronary calcium score CT scans to identify patients at elevated risk of heart failure events. That work, detailed in a Scientific Reports paper, focuses on calcified plaque and surrounding structures rather than epicardial fat. Although the inputs and validation criteria differ, both approaches share a core insight: routine CT images contain far more prognostic information than a human reader typically extracts.
Scale without extra scans
One practical advantage sets this tool apart from many experimental biomarkers. It does not require a new blood draw, a specialized imaging protocol, or an additional appointment. The Oxford Medical Sciences Division states that roughly 350,000 cardiac CT referrals take place in the UK each year, a figure this publication has not independently verified. If the AI were embedded in routine reporting workflows, every one of those scans could generate a heart failure risk score at no extra imaging cost to the patient.
Globally, heart failure affects an estimated 64 million people, according to figures cited by the World Heart Federation, and hospitalizations for the condition place enormous strain on health systems. Catching high-risk individuals years before symptoms emerge could open a window for preventive treatment with medications already proven to slow or prevent heart failure progression, including ACE inhibitors, beta-blockers, and the newer SGLT2 inhibitors. The question is whether that theoretical window translates into real clinical benefit.
What still needs to happen
A promising accuracy figure is not the same as a proven clinical tool, and several gaps remain before this AI model could appear on an NHS radiology workstation.
No prospective trial yet. The study is retrospective: the model was tested on scans and outcomes that had already occurred, without influencing patient care at the time. Prospective trials, where the tool’s predictions guide real-time clinical decisions and outcomes are tracked forward, have not been publicly announced. Regulatory approval, medical device certification, and workflow integration would each add further time.
Subgroup performance is undisclosed. The published reporting does not break down accuracy by ethnicity, age band, or sex. AI models trained on one demographic mix can lose reliability when applied to different populations. Until those analyses are available, it is unclear whether the tool performs equally well for all patient groups.
Cost-effectiveness is unquantified. Running the algorithm on existing images is likely inexpensive. The larger cost comes from acting on new risk information: more frequent follow-up appointments, echocardiograms, blood tests, and earlier medication. Whether the downstream savings from prevented hospitalizations outweigh those costs has not been modeled for this specific tool.
Clinician trust and interface design are untested. The published work focuses on statistical metrics like the C-statistic, sensitivity, and specificity. It offers less detail on how results would be presented in practice. A single risk percentage, a tiered category, or a combined score incorporating traditional factors such as blood pressure and diabetes status could each influence clinical behavior differently. Without usability studies and feedback from cardiologists and radiologists, adoption is uncertain.
Scanner evolution could affect accuracy over time. CT hardware and image reconstruction algorithms change as hospitals upgrade equipment. Subtle shifts in image texture could cause an algorithm trained on one generation of scanners to drift in performance. The multi-site design of the current study suggests some resilience to that variation, but long-term stability will require ongoing monitoring and periodic recalibration.
Where this fits in the evidence hierarchy
In clinical research terms, this study sits above the typical single-center pilot but below a randomized controlled trial. It is advanced observational work with external validation across multiple sites, which is a meaningful threshold. It does not yet prove that using the tool improves patient outcomes, only that the tool can identify who is at risk.
The convergence of evidence from independent teams, spanning epicardial fat radiomics, perivascular fat attenuation, and calcium score analysis, reduces the likelihood that the Oxford result is a statistical fluke. Multiple groups, publishing in different journals, are finding that CT-derived fat and plaque features carry real cardiovascular prognostic power. That pattern of replication matters.
What this means for patients with upcoming cardiac CT scans
For anyone who has a cardiac CT scheduled or recently completed, the practical impact is not yet direct. No hospital is currently generating AI-driven heart failure risk scores from routine scans outside a research setting. The model remains in the study phase, and its path to clinical deployment depends on trials, regulation, and health system decisions that have not been finalized.
But the broader signal is hard to ignore. Routine medical images, captured for one purpose, may contain layers of prognostic information that human readers were never trained to see. As tools like the Oxford model mature and, after proper validation, potentially enter everyday practice, a single cardiac CT could increasingly answer not just what is happening inside the heart today, but what is likely to happen to it five or ten years from now. For a condition that affects tens of millions of people and often arrives too late for the most effective interventions, that shift from reactive diagnosis to predictive screening could matter enormously.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
