A Blind Spot in the Gold Standard
If you have high blood pressure and see a European cardiologist, your ten-year risk of a heart attack, stroke or heart-failure admission is probably calculated with SCORE2, the risk chart endorsed by the 2023 European Society of Hypertension guidelines [1]. SCORE2 is elegantly simple: feed in age, sex, systolic blood pressure, cholesterol and smoking status, and out pops a colour-coded band [2]. Simplicity is its virtue and its problem. A five-variable snapshot cannot see how your blood pressure has drifted over five years, how many pills you now take to control it, or whether your kidneys have started to slip.
A team at the Hospital de Sagunto in Valencia decided to test just how much SCORE2 misses. They pulled the records of every hypertensive patient the clinic had seen between 1991 and 2023 — 8,365 people — and filtered down to 3,588 who had no cardiovascular disease at their first visit and at least a year of follow-up [3]. Over an average of 8.3 years, 498 of them went on to have a first cardiovascular event — heart attacks, strokes and heart-failure admissions, at a rate of 1.93 events per 100 patient-years.
Then came the uncomfortable audit. When the researchers re-ran SCORE2 on those 498 patients using data from their baseline visit, only 32% had been classified as high or very-high risk. The remaining two-thirds — people who really did go on to have a cardiac event — had looked reassuringly ordinary on the day they walked in. In a category of tool that decides whether you get intensified prevention, missing two out of three future events is a strikingly large blind spot.
What the AI Actually Did
The Sagunto team asked a straightforward question: could a machine-learning model, trained on the same clinic records, do better? They chose XGBoost, a gradient-boosting framework, for three properties the authors explicitly justify in the paper: it handles missing values in the data natively, it captures complex non-linear interactions between variables, and it copes well with high-dimensional clinical datasets [3]. Crucially, they did not restrict it to a single snapshot. From each patient's follow-up they engineered longitudinal features — minimum, maximum, mean, standard deviation and first-to-last change for every continuous variable — so the model could learn from trajectories, not just baseline points.
The full feature set started at 155 clinical variables and was pruned, using SHAP importance scores, down to 30. The model was trained on 70% of the cohort and evaluated on the untouched remaining 30%. In that internal validation set, the algorithm hit an area under the ROC curve of 0.856, with 81.3% sensitivity and 78.0% specificity — respectable numbers for a real-world electronic health record dataset. More telling is what happened when the same 498 patients-who-later-had-events were fed through both models: SCORE2 flagged 32%, XGBoost flagged 81% — a near-threefold jump in detection, using data every hypertension clinic already collects.
Some caveats come with those numbers. XGBoost had one advantage SCORE2 does not: it saw longitudinal follow-up, so it was effectively predicting an event using information accumulated up to the day before that event, while SCORE2 was forced to work from the baseline visit alone [3]. The authors are explicit that this is a pragmatic clinical benchmark, not a strict head-to-head. The positive predictive value of the AI was also modest — 30.7% — meaning most patients it flags will not actually have an event in the follow-up window. That is the price of a high-sensitivity screening tool, and it is the same trade-off mammograms and PSA tests make.
Why Pill Count Beat Blood Pressure
The most surprising result is what the AI decided mattered most. When the researchers ranked the 30 surviving variables by influence, the top slot went not to systolic pressure, cholesterol or age, but to the number of antihypertensive drugs a patient was on. The next three most influential variables were the use of blood-thinning medication, kidney function, and the lowest recorded LDL ("bad") cholesterol level. Actual blood-pressure readings sat further down, in a middle tier the authors describe as having only a "moderate positive impact" [3].
There is a plausible clinical reading of this. The number of drugs a patient needs is a running summary of how hard their hypertension is to control. Someone stable on one agent has a different disease burden from someone escalated to three or four. That escalation, tracked over years, quietly encodes information about vascular stiffness, kidney decline and end-organ damage that a single office reading cannot. It is a longitudinal biomarker hiding in plain sight on every prescription list. Similar signals — cumulative blood-pressure exposure and long-term BP trajectories — have been shown to sharpen risk prediction in other cohorts, but they are hard to calculate in a clinic and rarely used [4]. Pill count is a crude but readily available proxy.
This does not mean drug count causes heart attacks; the model is descriptive, not causal. But it does argue that the traditional focus on a single "present blood pressure" number, in isolation from a patient's treatment history, throws away a lot of information. It also lines up with a broader lesson from other machine-learning studies in hypertension: models built on longitudinal electronic records repeatedly outperform baseline-only risk equations [5].
From Clinic Record to Consumer Wrist
The Sagunto model is a prototype, not a product. It is single-centre, retrospective, has not yet been externally validated on other populations, and its positive predictive value of 30.7% means most people it flags will not go on to have an event in the follow-up window [3]. Those are meaningful caveats. But the direction it points in is worth taking seriously, and it is not unique: several other groups working with hypertensive cohorts have reported similar gains when longitudinal machine-learning replaces baseline-only regression [5].
The common thread is that the shape of a patient's history — how their pressure, kidney function, medications and lab values move over time — carries information that a five-variable snapshot throws away. That is precisely the kind of information consumer wearables now generate at high resolution: continuous heart-rate, activity, sleep, and in newer devices, cuffless blood-pressure estimates. Reviews of wearable cardiology have argued for several years that this data stream will only translate into clinical value once it is paired with algorithms designed to read trajectories rather than single points [6]. Studies like this one, still confined to hospital records, suggest what such algorithms may eventually look like — and hint that the working definition of "cardiovascular risk" is likely to become both more personal and more accurate as the two data sources meet.
References
- Mancia G, Kreutz R, Brunström M, et al. "2023 ESH Guidelines for the management of arterial hypertension", Journal of Hypertension 41 (2023) 1874-2071. https://doi.org/10.1097/HJH.0000000000003480
- SCORE2 working group and ESC Cardiovascular Risk Collaboration. "SCORE2 risk prediction algorithms: new models to estimate 10-year risk of cardiovascular disease in Europe", European Heart Journal 42 (2021) 2439-2454. https://doi.org/10.1093/eurheartj/ehab309
- Rodilla E, Arrizibita-Iriarte O, Miranda-Serrano B, et al. "Optimization of artificial intelligence models for prediction of new-onset cardiovascular disease in patients with arterial hypertension", PLOS Digital Health 5 (2026) e0001441. https://doi.org/10.1371/journal.pdig.0001441
- Pool LR, Ning H, Wilkins J, Lloyd-Jones DM, Allen NB. "Use of Long-term Cumulative Blood Pressure in Cardiovascular Risk Prediction Models", JAMA Cardiology 3 (2018) 1096-1100. https://doi.org/10.1001/jamacardio.2018.2763
- Feng Y, Leung AA, Lu X, Liang Z, Quan H, Walker RL. "Personalized prediction of incident hospitalization for cardiovascular disease in patients with hypertension using machine learning", BMC Medical Research Methodology 22 (2022) 325. https://doi.org/10.1186/s12874-022-01814-3
- Bayoumy K, Gaber M, Elshafeey A, et al. "Smart wearable devices in cardiovascular care: where we are and how to move forward", Nature Reviews Cardiology 18 (2021) 581-599. https://doi.org/10.1038/s41569-021-00522-7