"New AI blood test predicts heart disease 15 years early," ScienceDaily announced on 19 July. The study under that headline is careful work, and it contains a figure the coverage left out. Set against a questionnaire, a physical exam and an ordinary blood count, the whole package of 2,920 proteins, 168 metabolites and a DNA score improves the chance of telling a future heart patient from a healthy one by about three in a hundred. The proteins on their own buy about two.
The work is led from the University of Hong Kong by Qingpeng Zhang, though it is less HKU than the coverage suggests: four of the eight authors have an HKU tie and four do not, including co-corresponding author Tong Liu, who is in Tianjin. First author Yan Luo is at City University of Hong Kong as well as an HKU centre in Shanghai, and the others sit at Nanjing, the Chinese University of Hong Kong, Liverpool and Aalborg. Their framework, CardiOmicScore, trains two neural networks to read, from a single blood sample, 2,920 proteins and 168 metabolites, the small molecules left behind as a body burns fuel, and turn them into a risk score for each of six cardiovascular conditions. It ran in Nature Communications in February. The pitch, in the paper and louder outside it, is that this beats a score read from your DNA, because genes are fixed at birth while proteins move. The data come from the UK Biobank, a British cohort of half a million volunteers; the headline numbers rest on about 24,000 of them, followed for up to 16.6 years.
Every protein-score gain is genuine and none reaches four pairs in a hundred
Take one person who went on to develop coronary artery disease, the narrowing behind most heart attacks, and one who did not, and ask a model which is which. A questionnaire, a physical exam and an ordinary blood count already get that pair right about 73 times in 100. Add all 2,920 proteins, all 168 metabolites and a score built from the person's DNA, and it gets them right about 76 times in 100. Three more pairs in a hundred. (The measure is Harrell's C-index, moving from 0.730 to 0.758, where 0.5 is a coin toss.)
Strip out the metabolites and the DNA score, and the proteins alone buy less, unevenly. They do most for heart failure, 80 pairs in a hundred against 77, and least for stroke and coronary artery disease itself, the two the coverage leads with, where the gain is under two pairs in a hundred. All six protein-score improvements exclude zero, so all six are genuine, and none of the six reaches four. Put the metabolites and the DNA score back and the ceiling, across every combination the paper tries, is clots in the veins and lungs, at 4.9 pairs in a hundred. The DNA score fares worst of all: at ranking who will develop heart failure it does worse than age and sex alone.
Then the number a person would actually want. Among the third of participants whose protein score for coronary artery disease was highest, about 154 in every 1,000, or 15 percent, developed it within 15 years. Which means about 846 in every 1,000 of the people the model ranked most alarming did not, across those same 15 years. For peripheral artery disease, which the score reads best of all, the top third's 15-year rate was about 36 in 1,000. These are survival estimates rather than headcounts, because people leave a study and the follow-up ends. They come from a cohort 93.7 percent white, aged 40 to 69 at entry, and healthier than the country it was drawn from: only 5.5 percent of those invited joined, and those who did smoke less, drink less and die less than their contemporaries. Good ranking on an uncommon disease still mostly flags people who never get it.
A Gothenburg group got there from a different direction, though not from different people: only about 54,000 UK Biobank participants have proteomics at all, so their 38,380 and this study's 43,373 must overlap heavily. Everything else differs. They used 114 selected proteins rather than 2,920, judged on 10-year fatal and non-fatal heart attacks, strokes and artery-clearing procedures rather than 15-year single conditions. Added to the standard European risk equation, the proteins moved it from 74 to 77 on the same scale, where 50 is a coin toss, and the verdict went into their title in the European Journal of Preventive Cardiology: proteomics "modestly improves prediction." Different proteins, different endpoint, different team, the same size of answer.
The paper ranks, and the press release warns
The paper claims the scores "significantly enhance risk prediction across CVDs up to 15 years prior to disease onset when combined with clinical data." Read the methods and those 15 years turn out to be a forecasting horizon, the window the model was asked to put a probability on, not a moment of discovery. Nothing in the paper detects anything. It ranks. HKUMed's press release in March had the tool "accurately forecast the future risk" and "provide early warning signals up to 15 years before clinical onset," and called the molecules "real-time recorders" of the body. By July ScienceDaily had the model able to "detect warning signals as far as 15 years before clinical onset," crediting its material to the university. Neither "detect" nor "forecast" appears anywhere in the paper; I searched it. Zhang, quoted inside that same release, is careful in a way the prose around him is not. His tool, he says, "can potentially change the trajectory of disease through timely lifestyle modifications and early prevention."
The selling point running through all of it is that proteins move as your health does, so the score could update with you. It might. Nobody has checked. Every sample in the study was drawn once, at recruitment, between 2006 and 2010, and the authors list this as a limitation: a "static snapshot" that "may not capture the dynamic changes in a person's molecular profile over time." They also write that the signatures "are not clinically generalizable or applicable until they have undergone external validation," which has not yet happened, and that the platform is "relatively expensive and not yet widely accessible." Their own public calculator, which I opened, reports a 10-year risk, not a 15-year one, and asks you to enter your protein score as a percentile, which you cannot know without the panel.
No trial has tested what telling someone this number does
What a 15-year warning does to a healthy person is the question, and this paper cannot answer it. Nobody in the study was told their score, and it reports no harm of any kind because it looked for none. I could find no trial of what telling someone a proteomic risk score does, so the nearest evidence concerns other tests. In 2017 a Cochrane review pooled 41 trials and 194,035 people on systematically providing a conventional risk score, to patients, to clinicians, often both. Cholesterol fell slightly, blood pressure fell slightly, prescriptions rose. On cardiovascular events themselves the figure was 5.4 percent against 5.3 percent for usual care, but that rests on 3 of those trials and 99,070 people, the reviewers graded it low-quality, and two of the three reported afterward that they had been underpowered for exactly that endpoint. Which is why the verdict is that the effect is uncertain, not that there is none.
On harm there is one reassuring finding, from a different test again. A 2016 BMJ review of 18 trials of telling people their DNA-based disease risk found no effect on smoking, diet or exercise, and also "no adverse effects, such as depression and anxiety." But depression was measured in only two of those 18 trials and anxiety in three, on what the authors call "few data for these secondary outcomes." That is genetic risk besides, the very thing this test is pitched as improving on. So the honest summary is that nobody has shown a distant number does harm, and nobody has really looked.
The strongest argument the other way sits in the paper itself. It runs a decision-curve analysis, a way of weighing true alarms against false ones at different thresholds, and it comes out in the model's favor. What it concludes depends on where the alarm is set, and I have not tried to render its numbers in plain terms. You also cannot get this test: no approval, no regulator, no product, only a model trained on a research cohort. If a version of it does reach a clinic, it may well be worth having, the way a slightly better instrument is worth having.
What it will not do is tell you what happens to you. It moves you from a large group of people, most of whom stay well, into a smaller group, most of whom also stay well. The most accurate sentence written about this test so far is the authors' own: not clinically applicable until externally validated. That is not modesty. That is the finding.
