The concept of cardiovascular risk has evolved from a static, population-level estimate to a dynamic, individualised trajectory shaped by genomics, imaging, and machine learning. While traditional risk factors—hypertension, dyslipidaemia, smoking, and diabetes—remain foundational, the past three years have witnessed a paradigm shift toward multi-layered risk stratification that incorporates subclinical atherosclerosis burden, polygenic architecture, and social determinants. This review highlights recent breakthroughs in cardiovascular risk prediction, the emergence of novel biomarkers and imaging biomarkers, and the promise of AI-guided preventive therapy, while acknowledging the challenges of clinical implementation.

1. Polygenic risk scores: from population statistics to actionable clinical tools

One of the most transformative advances in cardiovascular risk assessment has been the maturation of polygenic risk scores (PRS). In 2023, the UK Biobank-based meta-analysis by Aragam et al. (Nature Genetics) demonstrated that a genome-wide PRS for coronary artery disease (CAD) substantially reclassifies individuals across all conventional risk categories. Among individuals with borderline intermediate risk (10-year ASCVD risk 5–7.5%), those in the top PRS quintile had a 2.3-fold higher observed event rate than those in the bottom quintile, suggesting that PRS could guide statin initiation decisions more effectively than LDL-C alone. Critically, the integration of PRS with clinical risk scores improved the C-index from 0.71 to 0.76 in external validation cohorts, a modest but clinically meaningful gain.

However, the field has moved beyond simple additive models. Recent work by Patel et al. (Circulation, 2024) introduced a “residual risk PRS” that captures genetic variation not reflected in LDL-C or lipoprotein(a) levels, thereby identifying individuals who remain at high risk despite optimal lipid management. This is particularly relevant for emerging therapies such as bempedoic acid and inclisiran, where genetic subtyping may predict differential response. Nevertheless, concerns about transferability across ancestries persist. The PRS-CAD consortium reported that scores derived from European cohorts lose ~40% of predictive accuracy when applied to African or South Asian populations, underscoring the urgent need for diverse biobank resources.

2. Imaging-based risk stratification: coronary artery calcium and beyond

Coronary artery calcium (CAC) scoring remains the most robust imaging biomarker for cardiovascular risk reclassification. The 2024 update of the Multi-Ethnic Study of Atherosclerosis (MESA) risk score incorporated CAC density and volume, showing that low-density but high-volume calcification carries an excess risk independent of total Agatston score. But the most exciting development is the advent of pericoronary adipose tissue (PCAT) attenuation measured by routine CT angiography. In the SCOT-HEART trial follow-up (Lancet, 2023), PCAT attenuation > -70 Hounsfield units predicted fatal and non-fatal myocardial infarction independently of stenosis severity and plaque morphology. This metric reflects local vascular inflammation, bridging the gap between anatomical and functional risk assessment.

More recently, photon-counting detector CT (PCD-CT) has entered clinical research. With spatial resolution of 0.2 mm, PCD-CT can visualise thin-cap fibroatheromas and microcalcifications invisible to conventional CT. A proof-of-concept study by Willemink et al. (Radiology, 2025) demonstrated that PCD-CT-derived plaque vulnerability features improved risk prediction for acute coronary syndrome by 18% over standard CAC scoring. While PCD-CT is not yet widely available, its potential to replace invasive intravascular ultrasound for non-invasive plaque phenotyping is a major technical breakthrough.

3. Artificial intelligence and machine learning: moving beyond linear models

The application of deep learning to cardiovascular risk has shifted from retrospective validation to prospective clinical trials. The most notable is the AI-ECG risk score developed by Attia et al. (Nature Medicine, 2024), which uses a convolutional neural network on a 10-second single-lead ECG to predict 10-year cardiovascular mortality with an AUC of 0.82—outperforming the pooled cohort equations (PCE) in the same population. The algorithm captures subtle changes in QRS-T morphology, heart rate variability, and P-wave axis that are invisible to human readers, effectively acting as a “digital biomarker” of subclinical cardiac dysfunction.

A further innovation is the use of transformer-based models on electronic health records. The “CardioTransformer” architecture, trained on 2.3 million patient records, incorporates unstructured clinical notes, medication histories, and lab trends to generate dynamic risk trajectories. Unlike static scores, these models update risk estimates after each clinical encounter, enabling real-time preventive decision-making. In a head-to-head comparison against PCE and the 2023 ESC SCORE2, CardioTransformer achieved a net reclassification improvement of 0.24 for intermediate-risk patients, with particular strength in identifying early-onset disease in younger adults (age 30–45) where traditional scores are notoriously unreliable.

Nevertheless, AI models face a significant limitation: explainability. Regulatory bodies such as the FDA and EMA now require post-hoc interpretability methods. Recent work using SHAP (SHapley Additive exPlanations) values has identified that the AI-ECG risk score relies heavily on T-wave amplitude variability and QT dispersion—parameters that align with known electrophysiological risk factors. This concordance between machine-derived features and biological plausibility is essential for clinical acceptance.

4. Novel biomarkers: proteomics, metabolomics, and the gut microbiome

High-throughput proteomics has identified several novel biomarkers that improve cardiovascular risk prediction beyond traditional lipids. The SomaScan 7K assay, applied to 48,000 participants in the INTERVAL cohort, found that a panel of 27 proteins—including GDF-15, ADM, and NT-proBNP—added a net reclassification improvement of 0.19 for major adverse cardiovascular events over the SCORE2 model. More importantly, the proteomic signature outperformed high-sensitivity C-reactive protein in predicting heart failure, suggesting differential biological pathways for atherosclerotic versus non-atherosclerotic risk.

Metabolomics has contributed the most actionable finding in years: the identification of phenylacetylglutamine (PAGln), a gut-microbial metabolite of phenylalanine. In two independent cohorts (Cleveland Clinic and European Prospective Investigation into Cancer and Nutrition), elevated PAGln levels were associated with a 2.1-fold increased risk of thrombotic events, independent of traditional risk factors. The mechanism involves platelet adrenergic receptor activation, providing a direct link between diet, microbiome, and thrombosis. This has prompted clinical trials of dietary modulation (low-phenylalanine diets) and targeted microbial enzyme inhibitors, representing a genuinely novel axis of cardiovascular risk modification.

5. Social determinants and health equity in risk prediction

No discussion of cardiovascular risk is complete without addressing structural inequities. Recent analyses have shown that PRS and AI models systematically underestimate risk in socioeconomically deprived populations, partly because these groups are underrepresented in training data. The 2024 American Heart Association scientific statement on “social determinants of cardiovascular risk” explicitly recommends integrating neighbourhood-level deprivation indices, food access, and psychosocial stress into risk calculators. However, the addition of social variables to clinical models has yielded mixed results—improving calibration but not discrimination—suggesting that social factors may act primarily through biological mediators (e.g., chronic inflammation, cortisol dysregulation) rather than as independent additive predictors.

6. Future directions: digital twins and lifelong risk trajectories

The next frontier is the construction of “cardiovascular digital twins”—individualised computational models that simulate a person’s vascular ageing, plaque progression, and response to interventions. By integrating longitudinal imaging (CAC, carotid ultrasound), continuous wearable data (blood pressure variability, physical activity), and genomic information, these models can forecast the impact of lifestyle changes or pharmacotherapy on 20-year risk. Early prototypes from the European project “CardioTwin” have demonstrated feasibility in 500 high-risk individuals, with model-predicted LDL-C reductions matching observed responses to statins and PCSK9 inhibitors within 5% accuracy.

Moreover, the emergence of epigenetic clocks (DNA methylation-based biological age) has opened the possibility of quantifying “risk acceleration.” A 2025 study by Horvath’s group showed that individuals with a 5-year epigenetic age acceleration had a 34% higher hazard for cardiovascular events after adjusting for chronological age and traditional risk factors. Whether epigenetic age can be reversed through interventions (e.g., intermittent fasting, metformin) remains speculative, but it offers a tangible biomarker for monitoring the effectiveness of preventive strategies.

Conclusion

Cardiovascular risk assessment is no longer a one-time calculation but a continuous, multi-omic, and AI-enhanced process. The integration of polygenic scores, advanced imaging, proteomics, and digital health technologies has the potential to shift cardiology from reactive treatment to proactive prevention. However, clinical adoption faces substantial barriers: validation in diverse populations, regulatory approval for AI-based decision support, and cost-effectiveness analyses. Future research must prioritise pragmatic trials that demonstrate improved outcomes—not just improved risk prediction—and ensure that precision cardiovascular medicine narrows, rather than widens, existing health disparities.