Blood proteins may expose hidden heart risk that standard scores miss

For decades, doctors have estimated a person’s chances of having a heart attack or stroke using risk equations built from age, sex, blood pressure, cholesterol, smoking and diabetes. Those tools, from the Framingham model to the Pooled Cohort Equations and newer models such as PREVENT, SCORE2, QRISK4 and China-PAR, have genuinely improved preventive care worldwide. Yet they rest on an uncomfortable fact: a large share of cardiovascular events occurs in people whom the equations judge to be at low risk. In the Prospective Urban Rural Epidemiology study, 41 percent of first major cardiovascular events happened in individuals classified as low risk by the estimated 10-year World Health Organization risk. Because preventive therapies such as statins are typically allocated according to predicted risk, these apparently low-risk people may be systematically excluded from interventions that could change the course of their disease.

That pattern, which the authors of a new Perspective in PLOS Medicine call the “low-risk paradox,” exposes a key limitation: equations built solely on traditional cardiovascular factors fail to measure a meaningful component of residual risk. Writing in the journal, Xun Tang and Pei Gao of Peking University argue that escaping this ceiling requires moving along two axes at once: what is measured, and what is predicted. Within the cardiovascular–kidney–metabolic framework, they note, decades of research mean that individual clinical measures now often add only a marginal improvement to the C-index, a measure of how well a model separates people who will develop an event from those who will not, typically less than 0.01 on a scale running from 0.5, no better than chance, to 1.0, perfect discrimination.

The first axis concerns the predictors fed into a model. Genome-wide polygenic risk scores reveal lifelong susceptibility and are well suited to early-life stratification and primordial prevention, but because genotypes are fixed at conception they cannot capture the dynamic state of disease. The plasma proteome offers a fundamentally different signal, capturing the real-time interplay of genome, environment and disease activity. Unlike genotype, it is dynamic and potentially modifiable, so it could in principle track both baseline risk and the response to intervention. Falling assay costs and high-throughput platforms now allow hundreds to thousands of proteins to be measured from a single blood draw, bringing proteomic scores within reach of large-scale use.

Targeted panels of myocardial stress, inflammation and metabolic proteins have already been linked to incident cardiovascular events, although in large studies they have added only modest discrimination to conventional scores. Alongside molecular and imaging predictors, the Perspective highlights dynamic monitoring as a complementary category. Wearable devices capture continuous, real-world physiology between clinic visits, letting risk be estimated repeatedly rather than from a single static measurement. Such signals gain practical value when fed into electronic health record systems that can act on them. The DREAM score, for example, repeatedly updates short-term cardiovascular risk from real-world electronic health record data to trigger automated early warnings, while the LINK-HF project showed that continuous wearable monitoring analytics could predict heart failure hospitalization.

The second axis concerns the outcomes being predicted, which are expanding in both directions along the natural history of disease. At the earlier end, the target is shifting from clinical events to subclinical atherosclerosis, which is common even among people at low conventional risk and whose progression, integrating cumulative exposures with individual susceptibility, independently forecasts future events. Coronary artery calcium scoring already demonstrates that detecting subclinical disease can reclassify risk and guide statin allocation, but it depends on imaging that is neither cheap nor universally available. At the later end, newer algorithms have begun to incorporate heart failure as a distinct outcome, reflecting its rising burden and the potential for prevention even in forms such as heart failure with preserved ejection fraction. PREVENT now estimates heart failure risk alongside atherosclerotic cardiovascular disease, and SCORE2-HF extends the original SCORE2 to predict heart failure in the general population.

Together, the two axes reframe cardiovascular risk prediction as managing a trajectory rather than forecasting a single event. The Perspective points to a recent study published in the same journal as occupying the intersection of both axes: pairing an omics predictor with an early outcome by linking the plasma proteome to the progression of subclinical atherosclerosis. From repeated coronary artery calcium and carotid plaque imaging in a Chinese community cohort, the investigators derived protein risk scores for each trait, then tested them against incident atherosclerotic cardiovascular disease in a separate case-cohort benchmarked against the China-PAR model. The idea is that a blood-based signature of that progression could deliver similar information more accessibly, flagging candidates for intensified prevention before irreversible injury occurs.

The distance between a statistical signal and a clinically useful tool, however, remains substantial, and the authors stress that this hurdle is general rather than specific to any single study. The protein risk score study is a case in point: its scores raised the C-index by only 0.018 over China-PAR, and on their own they discriminated incident atherosclerotic cardiovascular disease only a little better than age and sex. Their value lay instead in net reclassification improvement, which also proved double-edged, returning about one in five intermediate-risk future events to low risk. Scores derived and validated within a single population and platform must still prove themselves across ancestries and assays, and those trained to predict an imaging endpoint must also show value over directly measuring that endpoint. The study’s repeated-measures design, internally consistent signatures and acceptable calibration are, the authors write, the right foundation on which to build.

Whether acting on such signals can actually change outcomes is no longer hypothetical, according to the Perspective. In a pragmatic randomized trial, an AI-enabled electrocardiogram alert that flagged hospitalized patients at high risk of death prompted more timely intensive care and significantly lowered 90-day all-cause mortality. The authors argue that evidence showing acting on these scores prevents disease cost-effectively is still required, and that longitudinal designs testing whether within-person change tracks treatment response are also needed. Translation can also flow backward toward the laboratory: population-derived signatures, in this case pointing to cathepsin D and IL-2RA, suggest mechanisms and potential drug targets for further study. Machine-learning methods that extend regression-based models may help push prevention toward a more proactive, personalized paradigm.

The omics era, the authors conclude, will not abolish clinical risk scores; it will be judged by whether it extends their reach to the many patients those scores now overlook. The proteomic signatures examined in the companion study lie at the current frontier of cardiovascular risk prediction, making visible a risk that clinical factors cannot, while stopping short of external validation across populations. As the Perspective frames it, the reorientation is already under way: predictors are broadening from clinical factors to molecular signatures, while predicted outcomes are expanding earlier to subclinical atherosclerosis and later to heart failure. The question that will determine the field’s success is whether these molecular tools can move from statistical promise to demonstrated, cost-effective prevention for the people current equations leave behind.

Subject of Research: Reorienting cardiovascular risk prediction using proteomic signatures and expanded outcomes across the disease trajectory

Article Title: Seeing the invisible: Reorienting cardiovascular risk prediction across the disease trajectory

Article References: Tang, X., & Gao, P. (2026). Seeing the invisible: Reorienting cardiovascular risk prediction across the disease trajectory. PLOS Medicine, 23(10), e1005263. https://doi.org/10.1371/journal.pmed.1005263

Image Credits: AI Generated

DOI: 10.1371/journal.pmed.1005263

Keywords: cardiovascular risk prediction, proteomics, protein risk scores, subclinical atherosclerosis, heart failure, polygenic risk scores, coronary artery calcium, electronic health records, wearable devices, PREVENT, statins, risk reclassification