For most of medical history, a heart attack or a stroke has announced itself with brutal suddenness—a crushing chest pain, a collapsed limb, a hospital corridor at 3 a.m. Yet the biology that leads to those moments unfolds slowly, over years, in the quiet chemistry of the bloodstream. A research team at the LKS Faculty of Medicine of the University of Hong Kong (HKUMed) has now built a tool that tries to read that slow chemistry before it turns loud. Called CardiOmicScore, the artificial-intelligence system uses a single blood test to estimate a person's future risk of six major cardiovascular diseases, and in people already on a dangerous trajectory it flagged elevated risk as much as fifteen years before any symptoms appeared.
The six conditions the model was trained to anticipate read like a catalogue of the ways a circulatory system can fail: coronary artery disease, stroke, heart failure, atrial fibrillation, peripheral artery disease, and venous thromboembolism. What unites them is that each is far more treatable when caught early, and each is routinely missed until it is not. The findings were published in Nature Communications, and the study was led by Professor Zhang Qingpeng of HKUMed's Department of Pharmacology and Pharmacy together with the HKU Musketeers Foundation Institute of Data Science, with Luo Yan as first author.
The instinct to describe this as a crystal ball is worth resisting. CardiOmicScore is not predicting the future so much as it is listening more carefully to the present—translating thousands of faint molecular signals into a single, legible number that a clinician and a patient can actually act on.
Why It Matters
Cardiovascular disease remains the leading cause of death on the planet, responsible for roughly 19.8 million deaths in 2022 alone. The tools clinicians use to gauge who is at risk have changed surprisingly little in decades. A routine assessment still leans on age, blood pressure, cholesterol, smoking history, and a handful of other conventional markers. These measures are useful, but they are blunt: they describe the categories a person belongs to rather than the specific biological weather moving through their body right now.
Genetic risk scores were supposed to sharpen that picture. Polygenic risk scores bundle the effects of many DNA variants into a single estimate of inherited susceptibility, and they have grown popular over the past decade. But a genome is fixed at birth. It cannot register the toll of a decade of poor sleep, the benefit of a new exercise habit, the slow inflammation of an untreated condition, or the environmental exposures that accumulate across a working life. A genetic score tells you where you started. It says little about where you are.
CardiOmicScore is built to close precisely that gap. By reading the body's current molecular state rather than its inherited baseline, it aims to capture the dynamic, changeable component of risk—the part that lifestyle and medicine can still influence. That is a meaningful philosophical shift as much as a technical one: it reframes cardiovascular risk as something in motion rather than something sealed at conception.
How the Model Listens
The technical heart of the work is an approach called multiomics. Rather than examining a single layer of biology, the HKUMed team used deep learning to weave together three of them at once: genomics, which reads genetic information; proteomics, which measures the proteins that carry out most of the body's essential work; and metabolomics, which tracks the small molecules produced as the body burns fuel, fights infection, and responds to stress.
Working with large-scale population data from the UK Biobank, the model analysed 2,920 circulating proteins and 168 metabolites drawn from blood samples. Individually, any one of those molecules is a whisper. Taken together, they behave like a set of real-time recorders, registering subtle shifts in immune activity, metabolism, and vascular health long before those shifts harden into diagnosable disease. The role of the artificial-intelligence model is to find the faint patterns hidden across nearly three thousand simultaneous signals—a task that comfortably exceeds what any human analyst could hold in mind at once.
“Genes determine where we start,” Professor Zhang explained, adding that proteins and metabolites instead reflect a person's current physical health. The distinction is the whole argument of the paper compressed into a sentence: the model is designed to decode molecular signals that change, so that risk can be caught while there is still time to change it too.
The Reaction and the Caveats
Within the precision-medicine community, work like this lands on receptive ground, because it fits a direction the field has been moving for years—away from static, one-size-fits-many risk categories and toward dynamic, individualised assessment. The reported performance is the part that draws attention: CardiOmicScore substantially outperformed conventional polygenic risk scores, and its accuracy improved further when researchers layered in ordinary clinical information such as age and sex. A fifteen-year warning window, in the populations where it held, is the kind of lead time that could reshape when and how prevention begins.
Enthusiasm, though, deserves discipline. The analysis rests on the UK Biobank, a cohort that skews older, healthier, and more European than the world it hopes to serve; whether the same signals hold across more diverse populations is an open and important question. Measuring nearly three thousand proteins is not yet a routine or inexpensive procedure, and a research result is a long way from a test a family doctor can order on a Tuesday. And an early warning is only as valuable as the action it enables—a risk score that frightens without guiding is of little use. These are not reasons for doubt so much as the ordinary distance between a strong study and settled clinical practice.
What Comes Next
The trajectory the researchers describe is one where a small blood sample could eventually generate a comprehensive cardiovascular risk profile spanning several diseases at once, rather than the single-condition estimates that dominate today. That vision points toward validation in broader and more varied populations, prospective trials that test whether earlier warnings actually change outcomes, and the unglamorous engineering work of making high-dimensional molecular measurement cheap and reproducible enough for everyday care.
Professor Zhang framed the ambition in terms of a shift from reactive treatment to proactive prevention—identifying and heading off disease before it develops. It is a goal the whole of preventive cardiology shares, and one that better prediction alone cannot deliver; it also demands that patients and health systems respond to a warning years before anything feels wrong, which is its own behavioural and economic challenge. Still, moving the moment of insight fifteen years earlier changes what that response could look like.
Closing Thoughts
There is something quietly radical about a test that treats the bloodstream as a diary rather than a snapshot. For a century, medicine has largely waited for the body to break and then rushed to repair it. Tools like CardiOmicScore gesture at a different posture—one that reads the accumulating evidence of harm while it is still reversible, and treats a decade of advance notice as an asset rather than a curiosity.
Whether this particular model becomes a fixture of the clinic or a stepping stone to the one that does, the underlying idea feels durable: that the most useful thing artificial intelligence can do in medicine may not be to diagnose the disease in front of us, but to notice the one still forming, quietly, years away.
한글 요약
홍콩대학교 의과대학(HKUMed) 연구팀이 단 한 번의 혈액 검사로 여섯 가지 주요 심혈관 질환의 미래 위험을 예측하는 인공지능 도구 ‘카디오믹스코어(CardiOmicScore)’를 개발했다. 관상동맥질환, 뇌졸중, 심부전, 심방세동, 말초동맥질환, 정맥혈전색전증이 대상이며, 고위험군에서는 증상이 나타나기 최대 15년 전에 위험 신호를 포착했다. 연구 결과는 국제학술지 네이처 커뮤니케이션즈에 게재됐고, 장칭펑(Zhang Qingpeng) 교수가 이끌었다.
핵심은 ‘멀티오믹스’ 접근이다. 연구팀은 유전체(genomics)뿐 아니라 단백질(proteomics)과 대사물질(metabolomics)까지 통합해, UK 바이오뱅크 데이터에서 혈중 단백질 2,920종과 대사물질 168종을 딥러닝으로 분석했다. 태어날 때 고정되는 유전 위험 점수와 달리, 이들 분자 신호는 생활습관·환경·노화에 따라 실시간으로 변하는 ‘현재의 건강 상태’를 반영한다. 실제로 카디오믹스코어는 기존 다유전자 위험 점수(PRS)보다 예측 성능이 크게 앞섰고, 나이·성별 같은 임상 정보를 더하면 정확도가 한층 높아졌다.
다만 이 연구는 상대적으로 고령·유럽계 중심인 UK 바이오뱅크에 기반해 다양한 인구집단으로의 일반화 검증이 남아 있고, 수천 종의 단백질 측정이 아직 값비싸다는 한계도 분명하다. 그럼에도 심혈관 질환이 2022년 한 해에만 약 1,980만 명의 목숨을 앗아간 세계 최대 사망 원인이라는 점을 감안하면, 치료 중심에서 조기 예측·예방 중심으로 의료의 무게추를 옮기려는 이 시도의 방향성은 주목할 만하다. (참고: Nature Communications, HKUMed 보도자료, ScienceDaily)