Design: Prediction model development and external validation; Cox models with competing-risk adjustment
Population: 7,698 patients with established ASCVD and no prior heart failure for development; 240,741 across six external data sources for validation
Primary outcome: Ten-year and lifetime risk of heart failure hospitalisation or heart failure death
Effect: 1,031 events (13%) over median 11.2 years in development and 24,885 (10%) in validation; pooled C-statistic 0.696 (95% CI 0.674-0.717) with predicted risk matching observed incidence

Risk assessment in established atherosclerotic cardiovascular disease (ASCVD) predicts recurrent atherothrombotic events. It does not predict incident heart failure, which is a common and separate way for these patients to deteriorate. SMART2-HF fills that gap using only routinely available clinical variables, and is deliberately aligned with the SMART2 model already recommended for recurrent risk in the same population.

It was developed in 7,698 people with coronary, cerebrovascular or peripheral arterial disease or abdominal aortic aneurysm and no prior heart failure, using Cox models with age as the time scale and competing non-heart-failure mortality accounted for. Over a median 11.2 years, 13% developed heart failure. External validation covered 240,741 patients across six data sources — UK primary care records, HUNT3, SWEDEHEART, the ASCVD-Particles cohort, the Estonian Biobank and the REACH registry — with 24,885 events. The pooled C-statistic was 0.696 (95% CI 0.674-0.717), consistent across sex and ASCVD type, and predicted risk matched what actually happened.

What changes is the conversation. In a patient with vascular disease you already follow, you can now say what their ten-year and lifetime heart failure risk looks like on the numbers in front of you — which is the prerequisite for arguing about blood pressure control, weight, and the agents with heart failure benefit.

Use it in patients with established ASCVD and no prior heart failure — the population it was built and validated in
Only routinely collected variables are needed; no new test is required
Pair it with SMART2 rather than replacing it: the two answer different questions
Treat the output as a prompt for risk-factor intensification, not as an indication in itself
Note that no Indian validation cohort was included; calibration here is untested

Why it matters

Heart failure risk in vascular disease has been carried as an impression rather than a number, and impressions do not drive intensification.

The statistics, in plain English

A C-statistic of 0.696 means that given one patient who develops heart failure and one who does not, the model ranks them correctly about 70% of the time — useful for grouping, not for deciding an individual case. The more important property here is calibration: predicted risk matched observed incidence in the validation sets, so the percentage it produces can be quoted to a patient rather than merely used to sort them. Both were established outside the development cohort, which is what separates a usable model from an overfitted one.