Researchers at the University of Hong Kong’s medical school have built an AI system that reads a single blood sample and flags a person’s risk for six different cardiovascular diseases up to 15 years before symptoms appear. The tool, called CardiOmicScore, combines deep-learning models trained on large-scale biological data from the UK Biobank to predict coronary artery disease, stroke, heart failure, atrial fibrillation, peripheral artery disease and venous thromboembolism. The work, described in a Nature Communications paper, arrives at a moment when access to the very dataset that made it possible is restricted, raising questions about how quickly independent scientists can verify and extend the findings.
Why a 15-year cardiovascular warning window changes the calculus
Heart disease remains the leading cause of death globally, and most standard screening tools rely on a handful of clinical variables collected at a single office visit. CardiOmicScore takes a different approach. According to the HKUMed announcement, the system draws on two purpose-built neural networks, MetNet and ProNet, each trained on UK Biobank multiomics data that includes proteins, metabolites and genomic markers. By layering those biological signals on top of conventional risk factors, the models generate a personalized probability score for each of the six conditions.
The practical difference for patients is time. A 15-year lead on a heart failure or stroke diagnosis could shift treatment from emergency intervention to years of preventive management through statins, blood-pressure control, exercise or anticoagulation. That window is especially relevant for conditions like peripheral artery disease and venous thromboembolism, which often go undetected until a crisis forces a hospital visit. Earlier identification of high-risk individuals could also influence how aggressively clinicians manage coexisting conditions such as diabetes and chronic kidney disease, which compound cardiovascular risk.
From a health-system perspective, a long warning window changes resource planning. If a clinic can identify a subset of middle-aged patients with sharply elevated 15-year risk, it can prioritize them for intensive lifestyle programs, closer follow-up and earlier consideration of therapies such as PCSK9 inhibitors or GLP-1 receptor agonists. Conversely, patients with very low predicted risk might avoid unnecessary imaging or invasive testing. In principle, that risk stratification could make preventive cardiology more efficient, but only if the model proves reliable across diverse populations and practice settings.
A central question is whether a simpler, protein-only version of the score could perform nearly as well as the full multi-omics model. If independent teams gain access to UK Biobank proteomics data within the next year, they could test whether a stripped-down protein panel retains most of the predictive accuracy. A positive result would lower the cost and complexity of any future clinical test, making regulatory review more straightforward and bringing the technology closer to a routine blood draw at a primary-care office. It would also ease implementation in health systems that lack access to advanced metabolomics or whole-genome sequencing.
UK Biobank proteomics and the evidence trail behind CardiOmicScore
The Nature Communications study is not the first to show that blood proteins can forecast cardiovascular events over long time horizons. A separate analysis of UK Biobank proteomics data demonstrated that protein-based risk scores improve prediction beyond demographics and a broad set of lifestyle and clinical variables, with follow-up linkage extending up to roughly 15 years. That finding established the biological plausibility that CardiOmicScore now builds on with its deep-learning architecture.
Independent support comes from an Icelandic cohort study published in JAMA Cardiology, which tracked participants for a median of 15.8 years and found that proteomics-based risk scores predicted incident atherosclerotic cardiovascular events with meaningful discrimination gains over standard clinical models. The convergence of evidence from two distinct populations, one British and one Icelandic, strengthens the case that protein signals carry durable prognostic information across ethnic and geographic lines.
The underlying data pipeline matters, too. UK Biobank plasma samples are transferred to Olink for large-scale proteomic profiling, and the resulting data are returned to the biobank and its research consortium. That infrastructure, described on the pharma proteomics project page, has enabled a growing body of cardiovascular and dementia prediction research. A separate UK Biobank proteomics study published in Circulation extended the same approach to multiple cardiovascular outcomes and dementia, confirming that multi-disease prediction from blood proteins is a replicable finding rather than a one-off result.
CardiOmicScore fits into this trajectory by explicitly targeting six cardiovascular outcomes in a single framework. Instead of building one model for coronary artery disease and another for stroke, the HKU team trained MetNet and ProNet to capture shared and disease-specific signatures across conditions. That multi-task strategy could be especially valuable for patients whose risk profiles do not map neatly onto a single endpoint, such as those at simultaneous risk for atrial fibrillation and heart failure.
What stands between CardiOmicScore and a doctor’s office
Several gaps separate a promising research tool from a test that clinicians can order. The Nature Communications paper introduces the framework and its two neural networks, but specific performance metrics such as area-under-the-curve values, hazard ratios and calibration plots for each of the six diseases are available only in the full manuscript. Without those numbers circulating freely, other research groups cannot benchmark CardiOmicScore against existing risk calculators like the Pooled Cohort Equations used in U.S. cardiology guidelines.
Replication faces a structural barrier. UK Biobank’s data access application process has been paused while changes are made to its Research Analysis Platform. Until applications reopen, independent teams cannot run their own models on the same proteomics dataset. That delay matters because regulatory bodies in the United States and Europe typically require external validation before considering a diagnostic tool for clinical use. Every month the data remains gated is a month that validation studies cannot begin.
There are also questions of generalizability. UK Biobank participants skew healthier and less socioeconomically deprived than the general population, and most are of European ancestry. Before CardiOmicScore could be deployed widely, it would need testing in cohorts that better reflect the diversity of real-world patients, including those with limited access to care and higher baseline risk. Regulators and clinicians will want to see how the model performs when confronted with missing data, inconsistent follow-up and coexisting illnesses that are underrepresented in research volunteers.
Implementation logistics present another hurdle. Even a protein-only version of the score would require standardized assays, robust quality control and integration into electronic health records. Health systems would need to decide how often to repeat testing, how to communicate long-range risk estimates to patients and how to avoid overburdening clinics with follow-up for people flagged as high risk. Payers, meanwhile, would weigh the upfront cost of proteomic testing against potential savings from prevented heart attacks and strokes.
Finally, ethical and privacy considerations loom in the background. Multi-omics profiles can reveal information far beyond cardiovascular risk, raising concerns about data reuse and consent. If a model like CardiOmicScore becomes part of routine care, patients will need clear explanations of what is being measured, how predictions are generated and who can access the underlying data. Transparent reporting of model performance, limitations and update plans will be central to maintaining trust.
For now, CardiOmicScore sits at the intersection of cutting-edge biology and practical constraint. The scientific case for long-horizon cardiovascular prediction from blood proteins is increasingly strong, buttressed by converging evidence from UK Biobank and independent cohorts. Whether that promise translates into a test that reshapes preventive cardiology will depend on something more prosaic: reopening data access, enabling replication and building the regulatory and clinical infrastructure to turn a research algorithm into a tool that patients and doctors can actually use.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
