A bicycle on a dirt road feels trouble one pebble at a time, while a bullet train needs the rail to mutter its warnings miles ahead. In this week’s specimen jar we find a similar ambition: can a routine cardiac CT scan, already taken for one purpose, reveal early signs that the heart may later stumble into failure?

The paper, by Oikonomou and colleagues in the Journal of the American College of Cardiology, studies a most peculiar anatomical informant: epicardial adipose tissue, or EAT, the fat that sits directly on the heart. Not the distant, pantry-like fat of the waistcoat region, mind you. This is local fat, camped beside the myocardium with no polite fence between them, exchanging biochemical gossip like two neighbors with regrettably thin walls.

A Pocket of Fat With Opinions

Heart failure is what happens when the heart cannot pump or fill well enough to meet the body’s needs. It often arrives after years of quiet remodeling, inflammation, vascular strain, and metabolic mischief. The nuisance is that by the time symptoms appear, the plot has been thickening for ages.

EAT is biologically active. It responds to signals from nearby heart muscle and blood vessels. In Victorian field-naturalist terms, it is not merely padding. It is a damp little weather station strapped to the heart.

The researchers asked: what if CT images contain more information than doctors usually read from them? Radiomics is the art of extracting quantitative features from medical images - shapes, textures, intensities, spatial patterns - the sort of details the human eye ignores because it has, quite reasonably, other engagements. Ordinary image tools such as combb2.io may sharpen photographs for human viewing; radiomics is stranger and sterner, asking images to become measurable terrain.

The Machine Peers Into the Specimen

The study used coronary CT angiography scans from 72,751 adults across 9 UK centers, all without known heart failure or prior myocardial infarction at baseline. An automated pipeline segmented the epicardial fat and extracted 1,655 features. Then the investigators used a harmonized survival autoencoder - a phrase that sounds like a brass instrument built by a mathematician - to produce a fat radiomic profile for heart failure, or FRP-HF.

An autoencoder compresses complex data into a smaller representation. Add survival modeling, and the machine is not merely asking “what patterns exist?” but “which patterns foretell later trouble, and when?” That distinction matters. A model that finds decorative weirdness is a parlor trick. A model that predicts time-to-event risk may be a useful instrument, provided it behaves outside its home laboratory and does not faint when shown a new scanner.

The results are notable. In internal validation, 1,737 people developed heart failure over a median 5.1 years. In external validation, 363 did over 4.0 years. The FRP-HF score reached C-statistics of 0.869 internally and 0.850 externally. Each 25-percentile rise in the score was linked with roughly 4-fold higher adjusted heart failure risk, even after accounting for age, sex, conventional risk factors, coronary artery disease severity, and EAT volume. Those in the highest decile had nearly 20-fold higher risk than those in the lowest decile.

At this point one must resist shouting “behold!” and frightening the cardiology fellows.

Why This Is More Than Numerology in a Waistcoat

The clever bit is that EAT volume alone did not tell the full tale. Texture and composition mattered. This echoes recent “fat-omics” work showing that detailed epicardial fat features from CT calcium scoring can improve cardiovascular event prediction beyond simpler measures like volume or average Hounsfield units. Other 2025 work likewise found that epicardial fat and calcium features could help predict coronary arterial remodeling.

So the research joins a broader movement: using existing medical images opportunistically. The scan has already happened. The radiation has already been paid for, biologically and financially. If software can extract a second opinion from the same pixels, splendid. Waste not, want not, said the algorithm, wearing a tiny lab coat.

The Necessary Bucket of Cold Rainwater

Several cautions remain. This is risk prediction, not diagnosis. The model was trained on people undergoing CCTA, not the entire population strolling about with sandwiches and unresolved emails. External validation helps, but broader testing across countries, scanner vendors, care systems, and patient groups still matters. AI imaging models can drift over time, inherit bias, and produce numbers that look confident while concealing brittle assumptions. The 2025 multisociety statement on AI in cardiac CT and MRI makes this plain: deployment needs monitoring, fairness checks, workflow integration, and proof that the tool improves care rather than merely decorating reports.

The practical question is not “Can the model predict?” It is “What should clinicians do with that prediction?” Earlier prevention sounds lovely, but medicine is where lovely ideas go to submit paperwork. Future studies need to show whether acting on FRP-HF changes outcomes: more aggressive risk-factor control, closer follow-up, cardiometabolic treatment, or other targeted prevention.

Still, the central observation is deliciously odd: the fat around the heart may carry a radiomic fingerprint of future heart failure. The specimen twitches, the notebook opens, and the natural philosopher adjusts his spectacles.

References

1. Oikonomou EK, Chan K, Patel P, et al. “Early Prediction of Heart Failure From Routine Cardiac CT Using Radiomic Phenotyping of Epicardial Fat.” Journal of the American College of Cardiology, 2026. DOI: 10.1016/j.jacc.2026.02.5116. PMID: 41949519. PMCID: PMC13289806.

2. Hu T, Freeze J, Singh P, et al. “Artificial Intelligence Prediction of Cardiovascular Events Using Opportunistic Epicardial Adipose Tissue Assessments From Computed Tomography Calcium Score.” JACC: Advances, 2024. DOI: 10.1016/j.jacadv.2024.101188. PMID: 39372475.

3. Mastrodicasa D, van Assen M, Huisman M, et al. “Use of AI in Cardiac CT and MRI: A Scientific Statement.” Radiology, 2025. DOI: 10.1148/radiol.240516.

4. Baeßler B, Götz M, Antoniades C, Heidenreich JF, Leiner T, Beer M. “Artificial Intelligence in Coronary Computed Tomography Angiography.” Frontiers in Cardiovascular Medicine, 2023. DOI: 10.3389/fcvm.2023.1120361.

5. Lee J, Hu T, Williams MC, et al. “Detection of Arterial Remodeling Using Epicardial Adipose Tissue Assessment From CT Calcium Scoring Scan.” Frontiers in Cardiovascular Medicine, 2025. DOI: 10.3389/fcvm.2025.1543816.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.