A new analysis from the NOAH-AFNET 6 trial has shown that artificial intelligence (AI) can help unlock data from pacemakers to better understand stroke risk in patients with device-detected atrial fibrillation (AF). The study found that patients who spent more time in AF did not have a higher risk of stroke or other cardiovascular events than those who spent less time in AF. Researchers presented these findings as a late-breaking clinical trial at the 2026 Heart Rhythm Society (HRS) annual meeting (23–26 April, Chicago, USA).
While AF burden—the amount of time a patient spends in AF—has emerged as an important factor linked to stroke risk and outcomes, it remains difficult to capture at scale in clinical research and routine practice.
NOAH-AFNET 6 is an investigator-initiated trial evaluating anticoagulation in patients with device-detected AF. The main trial found that anticoagulants slightly reduced stroke risk but increased major bleeding, leading to early termination due to an unfavourable balance between benefits and risks.
The sub-study of NOAH-AFNET 6 presented at HRS 2026 shows that AI-driven models using natural language processing can automatically and reliably extract AF burden information from routine pacemaker reports, supporting more individualised risk assessment and potentially helping clinicians make more informed decisions about therapies like anticoagulation. Researchers applied this approach within NOAH–AFNET 6, analysing 11,964 pacemaker reports from 2,534 patients with device-detected AF to better understand how AF burden relates to clinical outcomes and response to the anticoagulant edoxaban.
“With AF on the rise worldwide, we need better ways to use existing clinical data to understand not only whether patients have AF, but how much of it they experience over time,” said Ulrich Schotten (Maastricht University, Maastricht, Netherlands). “This AI-driven large language model approach allows us to unlock meaningful insights from data we already collect and use them to advance research and improve patient care.”
Investigators found that the model identified AF burden—or ‘mode switch burden’—in more than 70% of reports, showing strong applicability to real-world data. In a validation sub-cohort, the large language model matched manual review in more than 98% of cases when AF burden data were available, and AF burden and mode switch burden showed high concordance, supporting the reliability of device-based measures. Baseline AF burden remained low across the study population. Over a median follow-up of 19 months, the magnitude of AF burden did not relate to thromboembolic risk, with low stroke rates even among patients with higher AF burden, and AF burden did not modify the effect of anticoagulation therapy either. Anticoagulation increased the risks of bleeding and death, regardless of AF burden, researchers report.
“These results highlight the need to better understand the effects of low-burden AF on stroke risk,” commented Paulus Kirchhof (UKE Hamburg, Hamburg, Germany). “This analysis provides a clear signal: low AF burden is associated with low stroke risk. Larger datasets will be needed to define which levels of AF burden warrant anticoagulation and which may instead require rhythm control strategies.”
