Atrial fibrillation is one of the most important preventable causes of ischemic stroke, yet it can remain undiagnosed for years—especially when episodes are intermittent or produce few symptoms. Screening can detect more AF than routine care, but monitoring everyone is expensive, inconvenient, and unlikely to be the most efficient use of healthcare resources.
A new international study tested whether electronic health records could help identify the people most likely to have newly detectable AF. The researchers developed a simplified machine-learning model called FIND-AF 2.0, validated it across five countries, and then prospectively tested whether it could guide short-term ECG monitoring.
The results are promising for improving screening efficiency. They do not yet prove that FIND-AF 2.0-guided screening prevents strokes or improves survival.
Why targeted screening matters
AF is an irregular heart rhythm that can allow blood to pool in the atria and form clots. If a clot travels to the brain, it can cause an ischemic stroke.
The condition may be asymptomatic, intermittent, or discovered only during evaluation for another illness. This creates a practical problem: a normal office ECG does not exclude paroxysmal AF, but prolonged monitoring of large populations requires substantial time and resources.
Current approaches to selecting people for AF screening often rely on age or the CHA₂DS₂-VASc score. The score estimates stroke risk in people with established AF and incorporates factors such as heart failure, hypertension, age, diabetes, prior stroke or transient ischemic attack, vascular disease, and sex. It can also help identify people who would be candidates for oral anticoagulation if AF were detected.
But these factors do not necessarily identify who is most likely to receive a new AF diagnosis in the immediate future. The researchers therefore asked whether routinely collected health-record data could offer more precise short-term risk stratification.
Building a simpler model
The original FIND-AF model, developed in UK primary-care records in 2022, used 75 variables. Although it performed well, that number made large-scale implementation difficult because many routinely collected variables are missing or inconsistently recorded.
FIND-AF 2.0 used 12 variables:
Age.
Sex.
Diabetes.
Heart failure.
Hypertension.
Previous stroke or transient ischemic attack.
Ischemic heart disease.
Chronic obstructive pulmonary disease.
Valvular heart disease.
Chronic kidney disease.
Rheumatoid arthritis.
Hyperthyroidism.
The model predicted the likelihood of a new diagnosis of AF or atrial flutter during the following six months. It was developed using UK primary-care data and externally validated using datasets from Japan, Israel, Canada, and Hong Kong.
The retrospective analyses included almost 13 million people across the five datasets. The model could be applied to all records in the external validation cohorts, an important practical advantage over risk scores that require measurements such as weight, height, smoking status, or blood pressure.
How well did it perform?
In the UK validation cohort, FIND-AF 2.0 had an area under the receiver operating characteristic curve of 0.819. Its performance was not significantly different from the original 75-variable FIND-AF model, which had an AUROC of 0.824.
In Japan, Canada, and China, FIND-AF 2.0 discriminated better between people who did and did not receive a new AF diagnosis than CHA₂DS₂-VASc and C2HEST. In Israel, all models performed well, with no statistically significant difference between them.
The most influential variables in the simplified model were age, heart failure, ischemic heart disease, valvular heart disease, and hypertension.
These results indicate that a smaller model can retain much of the predictive performance of the original model. They do not mean that the model can diagnose AF without ECG monitoring.
What the findings mean
The results suggest that FIND-AF 2.0 can concentrate monitoring among people more likely to have newly detectable AF. Its advantage is practical: it uses a relatively small number of variables commonly available in electronic records.
However, the model does not diagnose AF by itself. It identifies people who may benefit from ECG monitoring.
The study also did not prove that FIND-AF 2.0-guided screening prevents stroke. The prospective study had no randomized usual-care control group, and only 51 AF cases were detected. Intermittent ECG monitoring may also miss episodes occurring between recordings.
A separate analysis of 229,565 people with known AF found that those with high FIND-AF 2.0 risk had higher rates of ischemic stroke when they were not taking anticoagulants. This supports an association between the score and stroke risk, but it does not show that the model predicts stroke better than CHA₂DS₂-VASc or that screening based on it improves outcomes.
Why outcomes matter
AF screening remains debated. Some guidelines support ECG-based screening in selected older adults or people with additional stroke-risk factors, while the US Preventive Services Task Force has concluded that evidence remains insufficient to determine whether screening improves health outcomes.
The key question is not simply whether a strategy detects more AF. It is whether earlier detection and treatment reduce stroke and disability without causing excessive bleeding, anxiety, overdiagnosis, or unnecessary treatment.
That question requires randomized trials comparing FIND-AF 2.0-guided screening with routine care or other screening approaches.
Source: Circulation
