Researchers at the University of Oxford have achieved a groundbreaking advancement in cardiovascular medicine with an artificial intelligence (AI) tool that can predict the risk of heart failure up to five years before symptoms appear. Published in the prestigious Journal of the American College of Cardiology (JACC), the study leverages routine cardiac computed tomography (CT) angiography scans—commonly used to investigate chest pain—to identify subtle indicators of impending heart failure. This development positions Oxford at the forefront of AI-driven precision cardiology, potentially transforming how the UK's National Health Service (NHS) approaches preventive care for one of the nation's leading causes of hospital admissions and mortality.
🔬 The Science Behind the Oxford AI Predictor
The innovation centers on the analysis of epicardial adipose tissue (EAT), the fat layer directly surrounding the heart. EAT serves as a dynamic 'sensor' for myocardial stress, undergoing textural changes in response to early inflammatory signals from the heart muscle. These alterations, invisible to the naked eye on standard CT images, are captured through advanced radiomics—a technique that extracts quantitative features from medical images, including volumetric measurements, shape descriptors, and higher-order texture patterns derived from wavelet transformations.
The AI model, termed the Fat Radiomic Profile for Heart Failure (FRPHF), employs a harmonized survival autoencoder architecture. This deep learning framework processes 1,655 radiomic features from EAT, compressing them into a low-dimensional latent space before applying Cox proportional hazards regression to generate individualized absolute risk scores. Trained without human intervention, the system adjusts for confounders like age, sex, body mass index (BMI), hypertension, diabetes, chronic kidney disease, ischemic heart disease history, coronary artery disease (CAD) severity (via CAD-RADS score), and EAT volume itself.
Unprecedented Dataset from UK NHS Trusts
The study's robustness stems from its massive, real-world dataset: 72,751 adults without prior heart failure or myocardial infarction who underwent coronary CT angiography (CCTA) across nine NHS trusts in England between 2007 and 2022. The model was developed using 59,327 scans from seven centers and externally validated on 13,424 from two geographically distinct sites, with median follow-ups of 5.1 and 4.0 years, respectively. Patients were tracked via national registries like the National Institute for Cardiovascular Outcomes Research (NICOR), confirming incident heart failure through ICD-10 codes at least 30 days post-scan.
This opportunistic use of existing scans—performed on approximately 350,000 NHS patients annually for chest pain evaluation—highlights the tool's scalability. No additional imaging or costs are required, making it ideal for integration into routine radiology workflows.
Impressive Performance Metrics
FRPHF demonstrated superior discriminatory power, with concordance (C)-statistics of 0.869 (95% CI: 0.850-0.889) internally and 0.850 (95% CI: 0.831-0.870) externally. Each 25-percentile increase in the score correlated with nearly fourfold higher adjusted hazard ratios (HR 3.90 internal; 3.79 external; both P < 0.001). Strikingly, the top risk decile faced almost 20-fold greater risk compared to the bottom, with a one-in-four chance of heart failure within five years for high-risk individuals.
Adding FRPHF to conventional models significantly enhanced five-year prediction (ΔAUC 0.047; NRI 0.39), showing net clinical benefit on decision curve analysis. Associations held across demographics, ejection fraction spectra (HFrEF, HFmrEF, HFpEF), and subgroups, underscoring its broad applicability.
Leadership from Professor Charalambos Antoniades
Leading the effort is Professor Charalambos Antoniades, BHF Professor of Cardiovascular Medicine at Oxford's Radcliffe Department of Medicine. His Oxford Translational Cardiovascular Research Group has pioneered AI applications in imaging, including prior tools for coronary inflammation and 10-year cardiovascular risk prediction. Co-authors like Evangelos K. Oikonomou (Oxford/Yale) and collaborators from Queen Mary University London, University of Leicester, and NHS sites exemplify UK-wide academic-clinical synergy.
"We have used developments in bioscience and computing to take a big step forward in treating heart failure," Antoniades noted. Funded by the British Heart Foundation (BHF) and NIHR Oxford Biomedical Research Centre, this work builds on Oxford's legacy in cardiovascular AI.
Addressing the Heart Failure Crisis in the UK
Heart failure affects over 1 million people in the UK, with around 200,000 new diagnoses annually—one every three minutes. It accounts for one in six heart-related deaths (~110,000 yearly) and drives substantial NHS burden: up to 80% of diagnoses occur in hospitals, often late-stage. Modifiable risks like hypertension (30% adults), diabetes (5.8 million), and obesity contribute to 70% of CVD burden, including heart failure.
Dr. Sonya Babu-Narayan, BHF clinical director, praised the tool: "Heart failure is consistently diagnosed too late... This study demonstrates the power of harnessing technology to unlock improvements in cardiovascular care." Early intervention could extend healthy lifespans and ease hospital pressures.
Oxford's Vanguard in AI Cardiovascular Innovation
Oxford leads UK higher education in AI-cardiology fusion. Antoniades' team previously developed AI for plaque inflammation prediction and 10-year event risk from CT. Recent feats include radiotranscriptomics for 'virtual biopsies' and foundation models for heart health akin to ChatGPT. Collaborations with Yale, Cleveland Clinic, and UK peers amplify impact, supported by NIHR Oxford BRC.
This JACC publication elevates Oxford's global profile, fostering interdisciplinary talent in bioscience, computing, and medicine.
Path to NHS Integration and Challenges Ahead
Regulatory approval is sought for NHS rollout, potentially analyzing all 350,000 annual cardiac CTs. Upgrades target broader chest CTs (e.g., lung scans). Challenges include site harmonization (addressed via ComBat), generalizability beyond CCTA, and ethical AI deployment. Yet, decision curve analyses affirm clinical utility.
Career Opportunities in UK AI-Health Research
Oxford's success underscores demand for AI specialists in UK universities. Roles in computational biology, radiomics, and cardio-AI abound, blending medicine with data science. Programs like Oxford's DPhil in Medical Sciences train next-gen researchers amid BHF/NIHR funding surges.
Photo by Samuel Isaacs on Unsplash
Conclusion: A New Era for Preventive Cardiology
Oxford's AI predictor heralds precision prevention, empowering clinicians to stratify risks from routine scans. As UK universities drive such innovations, collaborations between academia, NHS, and funders promise healthier futures. Explore research careers shaping tomorrow's medicine.